rlawjdghek commited on
Commit
80ccb59
β€’
1 Parent(s): cfa4c0e

stableviton

Browse files
This view is limited to 50 files because it contains too many changes. Β  See raw diff
Files changed (50) hide show
  1. app.py +115 -68
  2. cldm/cldm.py +137 -0
  3. cldm/hack.py +111 -0
  4. cldm/model.py +9 -0
  5. cldm/plms_hacked.py +251 -0
  6. cldm/warping_cldm_network.py +357 -0
  7. configs/VITON512.yaml +100 -0
  8. ldm/data/__init__.py +0 -0
  9. ldm/data/util.py +24 -0
  10. ldm/models/autoencoder.py +202 -0
  11. ldm/models/diffusion/__init__.py +0 -0
  12. ldm/models/diffusion/ddim.py +377 -0
  13. ldm/models/diffusion/ddpm.py +1875 -0
  14. ldm/models/diffusion/dpm_solver/__init__.py +1 -0
  15. ldm/models/diffusion/dpm_solver/dpm_solver.py +1154 -0
  16. ldm/models/diffusion/dpm_solver/sampler.py +87 -0
  17. ldm/models/diffusion/plms.py +244 -0
  18. ldm/models/diffusion/sampling_util.py +22 -0
  19. ldm/modules/attention.py +330 -0
  20. ldm/modules/diffusionmodules/__init__.py +0 -0
  21. ldm/modules/diffusionmodules/model.py +852 -0
  22. ldm/modules/diffusionmodules/openaimodel.py +790 -0
  23. ldm/modules/diffusionmodules/upscaling.py +81 -0
  24. ldm/modules/diffusionmodules/util.py +271 -0
  25. ldm/modules/distributions/__init__.py +0 -0
  26. ldm/modules/distributions/distributions.py +92 -0
  27. ldm/modules/ema.py +80 -0
  28. ldm/modules/encoders/__init__.py +0 -0
  29. ldm/modules/encoders/modules.py +213 -0
  30. ldm/modules/image_degradation/__init__.py +2 -0
  31. ldm/modules/image_degradation/bsrgan.py +730 -0
  32. ldm/modules/image_degradation/bsrgan_light.py +651 -0
  33. ldm/modules/image_degradation/utils/test.png +0 -0
  34. ldm/modules/image_degradation/utils_image.py +916 -0
  35. ldm/modules/image_encoders/__init__.py +0 -0
  36. ldm/modules/image_encoders/modules.py +52 -0
  37. ldm/modules/image_encoders/xf.py +130 -0
  38. ldm/modules/midas/__init__.py +0 -0
  39. ldm/modules/midas/api.py +170 -0
  40. ldm/modules/midas/midas/__init__.py +0 -0
  41. ldm/modules/midas/midas/base_model.py +16 -0
  42. ldm/modules/midas/midas/blocks.py +342 -0
  43. ldm/modules/midas/midas/dpt_depth.py +109 -0
  44. ldm/modules/midas/midas/midas_net.py +76 -0
  45. ldm/modules/midas/midas/midas_net_custom.py +128 -0
  46. ldm/modules/midas/midas/transforms.py +234 -0
  47. ldm/modules/midas/midas/vit.py +491 -0
  48. ldm/modules/midas/utils.py +189 -0
  49. ldm/modules/util.py +130 -0
  50. ldm/util.py +237 -0
app.py CHANGED
@@ -1,13 +1,19 @@
 
1
  import os
2
  import sys
3
  import time
4
  from pathlib import Path
 
 
 
5
 
6
  import gradio as gr
7
- import torch
8
  from PIL import Image
 
9
 
10
- from utils_stableviton import get_mask_location
 
 
11
 
12
  PROJECT_ROOT = Path(__file__).absolute().parents[1].absolute()
13
  sys.path.insert(0, str(PROJECT_ROOT))
@@ -18,6 +24,8 @@ from preprocess.openpose.run_openpose import OpenPose
18
 
19
  os.environ['GRADIO_TEMP_DIR'] = './tmp' # TODO: turn off when final upload
20
 
 
 
21
 
22
  openpose_model_hd = OpenPose(0)
23
  parsing_model_hd = Parsing(0)
@@ -25,63 +33,110 @@ densepose_model_hd = DensePose4Gradio(
25
  cfg='preprocess/detectron2/projects/DensePose/configs/densepose_rcnn_R_50_FPN_s1x.yaml',
26
  model='https://dl.fbaipublicfiles.com/densepose/densepose_rcnn_R_50_FPN_s1x/165712039/model_final_162be9.pkl',
27
  )
28
- stable_viton_model_hd = ... # TODO: write down stable viton model
29
 
30
  category_dict = ['upperbody', 'lowerbody', 'dress']
31
  category_dict_utils = ['upper_body', 'lower_body', 'dresses']
32
 
33
- # import spaces # TODO: turn on when final upload
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
 
 
 
 
 
 
35
  # @spaces.GPU # TODO: turn on when final upload
36
-
37
-
38
- def process_hd(vton_img, garm_img, n_samples, n_steps, guidance_scale, seed):
39
  model_type = 'hd'
40
  category = 0 # 0:upperbody; 1:lowerbody; 2:dress
41
 
42
- with torch.no_grad():
43
- openpose_model_hd.preprocessor.body_estimation.model.to('cuda')
44
-
45
- stt = time.time()
46
- print('load images... ', end='')
47
- garm_img = Image.open(garm_img).resize((768, 1024))
48
- vton_img = Image.open(vton_img).resize((768, 1024))
49
- print('%.2fs' % (time.time() - stt))
50
-
51
- stt = time.time()
52
- print('get agnostic map... ', end='')
53
- keypoints = openpose_model_hd(vton_img.resize((384, 512)))
54
- model_parse, _ = parsing_model_hd(vton_img.resize((384, 512)))
55
- mask, mask_gray = get_mask_location(model_type, category_dict_utils[category], model_parse, keypoints)
56
- mask = mask.resize((768, 1024), Image.NEAREST)
57
- mask_gray = mask_gray.resize((768, 1024), Image.NEAREST)
58
- masked_vton_img = Image.composite(mask_gray, vton_img, mask) # agnostic map
59
- print('%.2fs' % (time.time() - stt))
60
-
61
- stt = time.time()
62
- print('get densepose... ', end='')
63
- vton_img = vton_img.resize((768, 1024)) # size for densepose
64
- densepose = densepose_model_hd.execute(vton_img) # densepose
65
- print('%.2fs' % (time.time() - stt))
66
-
67
- # # stable viton here
68
- # images = stable_viton_model_hd(
69
- # vton_img,
70
- # garm_img,
71
- # masked_vton_img,
72
- # densepose,
73
- # n_samples,
74
- # n_steps,
75
- # guidance_scale,
76
- # seed
77
- # )
78
-
79
- # return images
80
-
81
-
82
- example_path = os.path.join(os.path.dirname(__file__), 'examples')
83
- model_hd = os.path.join(example_path, 'model/model_1.png')
84
- garment_hd = os.path.join(example_path, 'garment/00055_00.jpg')
 
 
85
 
86
  with gr.Blocks(css='style.css') as demo:
87
  gr.HTML(
@@ -114,36 +169,28 @@ with gr.Blocks(css='style.css') as demo:
114
  gr.Markdown("## Experience virtual try-on with your own images!")
115
  with gr.Row():
116
  with gr.Column():
117
- vton_img = gr.Image(label="Model", type="filepath", height=384, value=model_hd)
118
  example = gr.Examples(
119
  inputs=vton_img,
120
  examples_per_page=14,
121
- examples=[
122
- os.path.join(example_path, 'model/model_1.png'), # TODO more our models
123
- os.path.join(example_path, 'model/model_2.png'),
124
- os.path.join(example_path, 'model/model_3.png'),
125
- ])
126
  with gr.Column():
127
- garm_img = gr.Image(label="Garment", type="filepath", height=384, value=garment_hd)
128
  example = gr.Examples(
129
  inputs=garm_img,
130
  examples_per_page=14,
131
- examples=[
132
- os.path.join(example_path, 'garment/00055_00.jpg'),
133
- os.path.join(example_path, 'garment/00126_00.jpg'),
134
- os.path.join(example_path, 'garment/00151_00.jpg'),
135
- ])
136
  with gr.Column():
137
  result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery", preview=True, scale=1)
138
  with gr.Column():
139
  run_button = gr.Button(value="Run")
140
  # TODO: change default values (important!)
141
- n_samples = gr.Slider(label="Images", minimum=1, maximum=4, value=1, step=1)
142
- n_steps = gr.Slider(label="Steps", minimum=20, maximum=40, value=20, step=1)
143
- guidance_scale = gr.Slider(label="Guidance scale", minimum=1.0, maximum=5.0, value=2.0, step=0.1)
144
- seed = gr.Slider(label="Seed", minimum=-1, maximum=2147483647, step=1, value=-1)
145
 
146
- ips = [vton_img, garm_img, n_samples, n_steps, guidance_scale, seed]
147
  run_button.click(fn=process_hd, inputs=ips, outputs=[result_gallery])
148
 
149
- demo.launch()
 
1
+ # import spaces # TODO: turn on when final upload
2
  import os
3
  import sys
4
  import time
5
  from pathlib import Path
6
+ from omegaconf import OmegaConf
7
+ from glob import glob
8
+ from os.path import join as opj
9
 
10
  import gradio as gr
 
11
  from PIL import Image
12
+ import torch
13
 
14
+ from utils_stableviton import get_mask_location, get_batch, tensor2img
15
+ from cldm.model import create_model
16
+ from cldm.plms_hacked import PLMSSampler
17
 
18
  PROJECT_ROOT = Path(__file__).absolute().parents[1].absolute()
19
  sys.path.insert(0, str(PROJECT_ROOT))
 
24
 
25
  os.environ['GRADIO_TEMP_DIR'] = './tmp' # TODO: turn off when final upload
26
 
27
+ IMG_H = 1024
28
+ IMG_W = 768
29
 
30
  openpose_model_hd = OpenPose(0)
31
  parsing_model_hd = Parsing(0)
 
33
  cfg='preprocess/detectron2/projects/DensePose/configs/densepose_rcnn_R_50_FPN_s1x.yaml',
34
  model='https://dl.fbaipublicfiles.com/densepose/densepose_rcnn_R_50_FPN_s1x/165712039/model_final_162be9.pkl',
35
  )
 
36
 
37
  category_dict = ['upperbody', 'lowerbody', 'dress']
38
  category_dict_utils = ['upper_body', 'lower_body', 'dresses']
39
 
40
+ # #### model init >>>>
41
+ config = OmegaConf.load("./configs/VITON512.yaml")
42
+ config.model.params.img_H = IMG_H
43
+ config.model.params.img_W = IMG_W
44
+ params = config.model.params
45
+
46
+ model = create_model(config_path=None, config=config)
47
+ model.load_state_dict(torch.load("./checkpoints/VITONHD_1024.ckpt", map_location="cpu")["state_dict"])
48
+ model = model.cuda()
49
+ model.eval()
50
+ sampler = PLMSSampler(model)
51
+ # #### model init <<<<
52
+ def stable_viton_model_hd(
53
+ batch,
54
+ n_steps,
55
+ ):
56
+ z, cond = model.get_input(batch, params.first_stage_key)
57
+ bs = z.shape[0]
58
+ c_crossattn = cond["c_crossattn"][0][:bs]
59
+ if c_crossattn.ndim == 4:
60
+ c_crossattn = model.get_learned_conditioning(c_crossattn)
61
+ cond["c_crossattn"] = [c_crossattn]
62
+ uc_cross = model.get_unconditional_conditioning(bs)
63
+ uc_full = {"c_concat": cond["c_concat"], "c_crossattn": [uc_cross]}
64
+ uc_full["first_stage_cond"] = cond["first_stage_cond"]
65
+ for k, v in batch.items():
66
+ if isinstance(v, torch.Tensor):
67
+ batch[k] = v.cuda()
68
+ sampler.model.batch = batch
69
+
70
+ ts = torch.full((1,), 999, device=z.device, dtype=torch.long)
71
+ start_code = model.q_sample(z, ts)
72
+
73
+ output, _, _ = sampler.sample(
74
+ n_steps,
75
+ bs,
76
+ (4, IMG_H//8, IMG_W//8),
77
+ cond,
78
+ x_T=start_code,
79
+ verbose=False,
80
+ eta=0.0,
81
+ unconditional_conditioning=uc_full,
82
+ )
83
 
84
+ output = model.decode_first_stage(output)
85
+ output = tensor2img(output)
86
+ pil_output = Image.fromarray(output)
87
+ return pil_output
88
+
89
  # @spaces.GPU # TODO: turn on when final upload
90
+ @torch.no_grad()
91
+ def process_hd(vton_img, garm_img, n_steps):
 
92
  model_type = 'hd'
93
  category = 0 # 0:upperbody; 1:lowerbody; 2:dress
94
 
95
+ openpose_model_hd.preprocessor.body_estimation.model.to('cuda')
96
+
97
+ stt = time.time()
98
+ print('load images... ', end='')
99
+ garm_img = Image.open(garm_img).resize((IMG_W, IMG_H))
100
+ vton_img = Image.open(vton_img).resize((IMG_W, IMG_H))
101
+ print('%.2fs' % (time.time() - stt))
102
+
103
+ stt = time.time()
104
+ print('get agnostic map... ', end='')
105
+ keypoints = openpose_model_hd(vton_img.resize((IMG_W, IMG_H)))
106
+ model_parse, _ = parsing_model_hd(vton_img.resize((IMG_W, IMG_H)))
107
+ mask, mask_gray = get_mask_location(model_type, category_dict_utils[category], model_parse, keypoints)
108
+ mask = mask.resize((IMG_W, IMG_H), Image.NEAREST)
109
+ mask_gray = mask_gray.resize((IMG_W, IMG_H), Image.NEAREST)
110
+ masked_vton_img = Image.composite(mask_gray, vton_img, mask) # agnostic map
111
+ print('%.2fs' % (time.time() - stt))
112
+
113
+ stt = time.time()
114
+ print('get densepose... ', end='')
115
+ vton_img = vton_img.resize((IMG_W, IMG_H)) # size for densepose
116
+ densepose = densepose_model_hd.execute(vton_img) # densepose
117
+ print('%.2fs' % (time.time() - stt))
118
+
119
+ batch = get_batch(
120
+ vton_img,
121
+ garm_img,
122
+ densepose,
123
+ masked_vton_img,
124
+ mask,
125
+ IMG_H,
126
+ IMG_W
127
+ )
128
+
129
+ sample = stable_viton_model_hd(
130
+ batch,
131
+ n_steps
132
+ )
133
+ breakpoint()
134
+ return sample
135
+
136
+
137
+ example_path = opj(os.path.dirname(__file__), 'examples')
138
+ example_model_ps = sorted(glob(opj(example_path, "model/*")))
139
+ example_garment_ps = sorted(glob(opj(example_path, "garment/*")))
140
 
141
  with gr.Blocks(css='style.css') as demo:
142
  gr.HTML(
 
169
  gr.Markdown("## Experience virtual try-on with your own images!")
170
  with gr.Row():
171
  with gr.Column():
172
+ vton_img = gr.Image(label="Model", type="filepath", height=384, value=example_model_ps[0])
173
  example = gr.Examples(
174
  inputs=vton_img,
175
  examples_per_page=14,
176
+ examples=example_model_ps)
 
 
 
 
177
  with gr.Column():
178
+ garm_img = gr.Image(label="Garment", type="filepath", height=384, value=example_garment_ps[0])
179
  example = gr.Examples(
180
  inputs=garm_img,
181
  examples_per_page=14,
182
+ examples=example_garment_ps)
 
 
 
 
183
  with gr.Column():
184
  result_gallery = gr.Gallery(label='Output', show_label=False, elem_id="gallery", preview=True, scale=1)
185
  with gr.Column():
186
  run_button = gr.Button(value="Run")
187
  # TODO: change default values (important!)
188
+ # n_samples = gr.Slider(label="Images", minimum=1, maximum=4, value=1, step=1)
189
+ n_steps = gr.Slider(label="Steps", minimum=20, maximum=100, value=50, step=1)
190
+ # guidance_scale = gr.Slider(label="Guidance scale", minimum=1.0, maximum=5.0, value=2.0, step=0.1)
191
+ # seed = gr.Slider(label="Seed", minimum=-1, maximum=2147483647, step=1, value=-1)
192
 
193
+ ips = [vton_img, garm_img, n_steps]
194
  run_button.click(fn=process_hd, inputs=ips, outputs=[result_gallery])
195
 
196
+ demo.queue().launch()
cldm/cldm.py ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from os.path import join as opj
3
+ import omegaconf
4
+
5
+ import cv2
6
+ import einops
7
+ import torch
8
+ import torch as th
9
+ import torch.nn as nn
10
+ import torchvision.transforms as T
11
+ import torch.nn.functional as F
12
+ import numpy as np
13
+
14
+ from ldm.models.diffusion.ddpm import LatentDiffusion
15
+ from ldm.util import instantiate_from_config
16
+
17
+ class ControlLDM(LatentDiffusion):
18
+ def __init__(
19
+ self,
20
+ control_stage_config,
21
+ validation_config,
22
+ control_key,
23
+ only_mid_control,
24
+ use_VAEdownsample=False,
25
+ config_name="",
26
+ control_scales=None,
27
+ use_pbe_weight=False,
28
+ u_cond_percent=0.0,
29
+ img_H=512,
30
+ img_W=384,
31
+ always_learnable_param=False,
32
+ *args,
33
+ **kwargs
34
+ ):
35
+ self.control_stage_config = control_stage_config
36
+ self.use_pbe_weight = use_pbe_weight
37
+ self.u_cond_percent = u_cond_percent
38
+ self.img_H = img_H
39
+ self.img_W = img_W
40
+ self.config_name = config_name
41
+ self.always_learnable_param = always_learnable_param
42
+ super().__init__(*args, **kwargs)
43
+ control_stage_config.params["use_VAEdownsample"] = use_VAEdownsample
44
+ self.control_model = instantiate_from_config(control_stage_config)
45
+ self.control_key = control_key
46
+ self.only_mid_control = only_mid_control
47
+ if control_scales is None:
48
+ self.control_scales = [1.0] * 13
49
+ else:
50
+ self.control_scales = control_scales
51
+ self.first_stage_key_cond = kwargs.get("first_stage_key_cond", None)
52
+ self.valid_config = validation_config
53
+ self.use_VAEDownsample = use_VAEdownsample
54
+ @torch.no_grad()
55
+ def get_input(self, batch, k, bs=None, *args, **kwargs):
56
+ x, c = super().get_input(batch, self.first_stage_key, *args, **kwargs)
57
+ if isinstance(self.control_key, omegaconf.listconfig.ListConfig):
58
+ control_lst = []
59
+ for key in self.control_key:
60
+ control = batch[key]
61
+ if bs is not None:
62
+ control = control[:bs]
63
+ control = control.to(self.device)
64
+ control = einops.rearrange(control, 'b h w c -> b c h w')
65
+ control = control.to(memory_format=torch.contiguous_format).float()
66
+ control_lst.append(control)
67
+ control = control_lst
68
+ else:
69
+ control = batch[self.control_key]
70
+ if bs is not None:
71
+ control = control[:bs]
72
+ control = control.to(self.device)
73
+ control = einops.rearrange(control, 'b h w c -> b c h w')
74
+ control = control.to(memory_format=torch.contiguous_format).float()
75
+ control = [control]
76
+ cond_dict = dict(c_crossattn=[c], c_concat=control)
77
+ if self.first_stage_key_cond is not None:
78
+ first_stage_cond = []
79
+ for key in self.first_stage_key_cond:
80
+ if not "mask" in key:
81
+ cond, _ = super().get_input(batch, key, *args, **kwargs)
82
+ else:
83
+ cond, _ = super().get_input(batch, key, no_latent=True, *args, **kwargs)
84
+ first_stage_cond.append(cond)
85
+ first_stage_cond = torch.cat(first_stage_cond, dim=1)
86
+ cond_dict["first_stage_cond"] = first_stage_cond
87
+ return x, cond_dict
88
+
89
+ def apply_model(self, x_noisy, t, cond, *args, **kwargs):
90
+ assert isinstance(cond, dict)
91
+
92
+ diffusion_model = self.model.diffusion_model
93
+ cond_txt = torch.cat(cond["c_crossattn"], 1)
94
+ if self.proj_out is not None:
95
+ if cond_txt.shape[-1] == 1024:
96
+ cond_txt = self.proj_out(cond_txt) # [BS x 1 x 768]
97
+ if self.always_learnable_param:
98
+ cond_txt = self.get_unconditional_conditioning(cond_txt.shape[0])
99
+
100
+ if cond['c_concat'] is None:
101
+ eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=None, only_mid_control=self.only_mid_control)
102
+ else:
103
+ if "first_stage_cond" in cond:
104
+ x_noisy = torch.cat([x_noisy, cond["first_stage_cond"]], dim=1)
105
+ if not self.use_VAEDownsample:
106
+ hint = cond["c_concat"]
107
+ else:
108
+ hint = []
109
+ for h in cond["c_concat"]:
110
+ if h.shape[2] == self.img_H and h.shape[3] == self.img_W:
111
+ h = self.encode_first_stage(h)
112
+ h = self.get_first_stage_encoding(h).detach()
113
+ hint.append(h)
114
+ hint = torch.cat(hint, dim=1)
115
+ control, _ = self.control_model(x=x_noisy, hint=hint, timesteps=t, context=cond_txt, only_mid_control=self.only_mid_control)
116
+ if len(control) == len(self.control_scales):
117
+ control = [c * scale for c, scale in zip(control, self.control_scales)]
118
+
119
+ eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=control, only_mid_control=self.only_mid_control)
120
+ return eps, None
121
+ @torch.no_grad()
122
+ def get_unconditional_conditioning(self, N):
123
+ if not self.kwargs["use_imageCLIP"]:
124
+ return self.get_learned_conditioning([""] * N)
125
+ else:
126
+ return self.learnable_vector.repeat(N,1,1)
127
+ def low_vram_shift(self, is_diffusing):
128
+ if is_diffusing:
129
+ self.model = self.model.cuda()
130
+ self.control_model = self.control_model.cuda()
131
+ self.first_stage_model = self.first_stage_model.cpu()
132
+ self.cond_stage_model = self.cond_stage_model.cpu()
133
+ else:
134
+ self.model = self.model.cpu()
135
+ self.control_model = self.control_model.cpu()
136
+ self.first_stage_model = self.first_stage_model.cuda()
137
+ self.cond_stage_model = self.cond_stage_model.cuda()
cldm/hack.py ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import einops
3
+
4
+ import ldm.modules.encoders.modules
5
+ import ldm.modules.attention
6
+
7
+ from transformers import logging
8
+ from ldm.modules.attention import default
9
+
10
+
11
+ def disable_verbosity():
12
+ logging.set_verbosity_error()
13
+ print('logging improved.')
14
+ return
15
+
16
+
17
+ def enable_sliced_attention():
18
+ ldm.modules.attention.CrossAttention.forward = _hacked_sliced_attentin_forward
19
+ print('Enabled sliced_attention.')
20
+ return
21
+
22
+
23
+ def hack_everything(clip_skip=0):
24
+ disable_verbosity()
25
+ ldm.modules.encoders.modules.FrozenCLIPEmbedder.forward = _hacked_clip_forward
26
+ ldm.modules.encoders.modules.FrozenCLIPEmbedder.clip_skip = clip_skip
27
+ print('Enabled clip hacks.')
28
+ return
29
+
30
+
31
+ # Written by Lvmin
32
+ def _hacked_clip_forward(self, text):
33
+ PAD = self.tokenizer.pad_token_id
34
+ EOS = self.tokenizer.eos_token_id
35
+ BOS = self.tokenizer.bos_token_id
36
+
37
+ def tokenize(t):
38
+ return self.tokenizer(t, truncation=False, add_special_tokens=False)["input_ids"]
39
+
40
+ def transformer_encode(t):
41
+ if self.clip_skip > 1:
42
+ rt = self.transformer(input_ids=t, output_hidden_states=True)
43
+ return self.transformer.text_model.final_layer_norm(rt.hidden_states[-self.clip_skip])
44
+ else:
45
+ return self.transformer(input_ids=t, output_hidden_states=False).last_hidden_state
46
+
47
+ def split(x):
48
+ return x[75 * 0: 75 * 1], x[75 * 1: 75 * 2], x[75 * 2: 75 * 3]
49
+
50
+ def pad(x, p, i):
51
+ return x[:i] if len(x) >= i else x + [p] * (i - len(x))
52
+
53
+ raw_tokens_list = tokenize(text)
54
+ tokens_list = []
55
+
56
+ for raw_tokens in raw_tokens_list:
57
+ raw_tokens_123 = split(raw_tokens)
58
+ raw_tokens_123 = [[BOS] + raw_tokens_i + [EOS] for raw_tokens_i in raw_tokens_123]
59
+ raw_tokens_123 = [pad(raw_tokens_i, PAD, 77) for raw_tokens_i in raw_tokens_123]
60
+ tokens_list.append(raw_tokens_123)
61
+
62
+ tokens_list = torch.IntTensor(tokens_list).to(self.device)
63
+
64
+ feed = einops.rearrange(tokens_list, 'b f i -> (b f) i')
65
+ y = transformer_encode(feed)
66
+ z = einops.rearrange(y, '(b f) i c -> b (f i) c', f=3)
67
+
68
+ return z
69
+
70
+
71
+ # Stolen from https://github.com/basujindal/stable-diffusion/blob/main/optimizedSD/splitAttention.py
72
+ def _hacked_sliced_attentin_forward(self, x, context=None, mask=None):
73
+ h = self.heads
74
+
75
+ q = self.to_q(x)
76
+ context = default(context, x)
77
+ k = self.to_k(context)
78
+ v = self.to_v(context)
79
+ del context, x
80
+
81
+ q, k, v = map(lambda t: einops.rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
82
+
83
+ limit = k.shape[0]
84
+ att_step = 1
85
+ q_chunks = list(torch.tensor_split(q, limit // att_step, dim=0))
86
+ k_chunks = list(torch.tensor_split(k, limit // att_step, dim=0))
87
+ v_chunks = list(torch.tensor_split(v, limit // att_step, dim=0))
88
+
89
+ q_chunks.reverse()
90
+ k_chunks.reverse()
91
+ v_chunks.reverse()
92
+ sim = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device)
93
+ del k, q, v
94
+ for i in range(0, limit, att_step):
95
+ q_buffer = q_chunks.pop()
96
+ k_buffer = k_chunks.pop()
97
+ v_buffer = v_chunks.pop()
98
+ sim_buffer = torch.einsum('b i d, b j d -> b i j', q_buffer, k_buffer) * self.scale
99
+
100
+ del k_buffer, q_buffer
101
+ # attention, what we cannot get enough of, by chunks
102
+
103
+ sim_buffer = sim_buffer.softmax(dim=-1)
104
+
105
+ sim_buffer = torch.einsum('b i j, b j d -> b i d', sim_buffer, v_buffer)
106
+ del v_buffer
107
+ sim[i:i + att_step, :, :] = sim_buffer
108
+
109
+ del sim_buffer
110
+ sim = einops.rearrange(sim, '(b h) n d -> b n (h d)', h=h)
111
+ return self.to_out(sim)
cldm/model.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ from ldm.util import instantiate_from_config
2
+
3
+
4
+ def get_state_dict(d):
5
+ return d.get('state_dict', d)
6
+
7
+ def create_model(config, **kwargs):
8
+ model = instantiate_from_config(config.model).cpu()
9
+ return model
cldm/plms_hacked.py ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """SAMPLING ONLY."""
2
+
3
+ import torch
4
+ import numpy as np
5
+ from tqdm import tqdm
6
+
7
+ from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor
8
+ from ldm.models.diffusion.sampling_util import norm_thresholding
9
+
10
+
11
+ class PLMSSampler(object):
12
+ def __init__(self, model, schedule="linear", **kwargs):
13
+ super().__init__()
14
+ self.model = model
15
+ self.ddpm_num_timesteps = model.num_timesteps
16
+ self.schedule = schedule
17
+
18
+ def register_buffer(self, name, attr):
19
+ if type(attr) == torch.Tensor:
20
+ if attr.device != torch.device("cuda"):
21
+ attr = attr.to(torch.device("cuda"))
22
+ setattr(self, name, attr)
23
+
24
+ def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
25
+ if ddim_eta != 0:
26
+ raise ValueError('ddim_eta must be 0 for PLMS')
27
+ self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
28
+ num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
29
+ alphas_cumprod = self.model.alphas_cumprod
30
+ assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
31
+ to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
32
+
33
+ self.register_buffer('betas', to_torch(self.model.betas))
34
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
35
+ self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
36
+
37
+ # calculations for diffusion q(x_t | x_{t-1}) and others
38
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
39
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
40
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
41
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
42
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
43
+
44
+ # ddim sampling parameters
45
+ ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
46
+ ddim_timesteps=self.ddim_timesteps,
47
+ eta=ddim_eta,verbose=verbose)
48
+ self.register_buffer('ddim_sigmas', ddim_sigmas)
49
+ self.register_buffer('ddim_alphas', ddim_alphas)
50
+ self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
51
+ self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
52
+ sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
53
+ (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
54
+ 1 - self.alphas_cumprod / self.alphas_cumprod_prev))
55
+ self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
56
+
57
+ @torch.no_grad()
58
+ def sample(self,
59
+ S,
60
+ batch_size,
61
+ shape,
62
+ conditioning=None,
63
+ callback=None,
64
+ img_callback=None,
65
+ quantize_x0=False,
66
+ eta=0.,
67
+ mask=None,
68
+ x0=None,
69
+ temperature=1.,
70
+ noise_dropout=0.,
71
+ score_corrector=None,
72
+ corrector_kwargs=None,
73
+ verbose=True,
74
+ x_T=None,
75
+ log_every_t=100,
76
+ unconditional_guidance_scale=5.,
77
+ unconditional_conditioning=None,
78
+ # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
79
+ dynamic_threshold=None,
80
+ **kwargs
81
+ ):
82
+ if conditioning is not None:
83
+ if isinstance(conditioning, dict):
84
+ ctmp = conditioning[list(conditioning.keys())[0]]
85
+ while isinstance(ctmp, list): ctmp = ctmp[0]
86
+ cbs = ctmp.shape[0]
87
+ if cbs != batch_size:
88
+ print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
89
+ else:
90
+ if conditioning.shape[0] != batch_size:
91
+ print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
92
+
93
+ self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
94
+ # sampling
95
+ C, H, W = shape
96
+ size = (batch_size, C, H, W)
97
+ print(f'Data shape for PLMS sampling is {size}')
98
+
99
+ samples, intermediates, cond_output_dict = self.plms_sampling(conditioning, size,
100
+ callback=callback,
101
+ img_callback=img_callback,
102
+ quantize_denoised=quantize_x0,
103
+ mask=mask, x0=x0,
104
+ ddim_use_original_steps=False,
105
+ noise_dropout=noise_dropout,
106
+ temperature=temperature,
107
+ score_corrector=score_corrector,
108
+ corrector_kwargs=corrector_kwargs,
109
+ x_T=x_T,
110
+ log_every_t=log_every_t,
111
+ unconditional_guidance_scale=unconditional_guidance_scale,
112
+ unconditional_conditioning=unconditional_conditioning,
113
+ dynamic_threshold=dynamic_threshold,
114
+ )
115
+ return samples, intermediates, cond_output_dict
116
+
117
+ @torch.no_grad()
118
+ def plms_sampling(self, cond, shape,
119
+ x_T=None, ddim_use_original_steps=False,
120
+ callback=None, timesteps=None, quantize_denoised=False,
121
+ mask=None, x0=None, img_callback=None, log_every_t=100,
122
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
123
+ unconditional_guidance_scale=1., unconditional_conditioning=None,
124
+ dynamic_threshold=None):
125
+ device = self.model.betas.device
126
+ b = shape[0]
127
+ if x_T is None:
128
+ img = torch.randn(shape, device=device)
129
+ else:
130
+ img = x_T
131
+
132
+ if timesteps is None:
133
+ timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
134
+ elif timesteps is not None and not ddim_use_original_steps:
135
+ subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
136
+ timesteps = self.ddim_timesteps[:subset_end]
137
+
138
+ intermediates = {'x_inter': [img], 'pred_x0': [img]}
139
+ time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
140
+ total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
141
+ print(f"Running PLMS Sampling with {total_steps} timesteps")
142
+
143
+ iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
144
+ old_eps = []
145
+
146
+ for i, step in enumerate(iterator):
147
+ index = total_steps - i - 1
148
+ ts = torch.full((b,), step, device=device, dtype=torch.long)
149
+ ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
150
+
151
+ if mask is not None:
152
+ assert x0 is not None
153
+ if i < self.first_n_repaint:
154
+ img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
155
+ img = img_orig * mask + (1. - mask) * img
156
+
157
+ outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
158
+ quantize_denoised=quantize_denoised, temperature=temperature,
159
+ noise_dropout=noise_dropout, score_corrector=score_corrector,
160
+ corrector_kwargs=corrector_kwargs,
161
+ unconditional_guidance_scale=unconditional_guidance_scale,
162
+ unconditional_conditioning=unconditional_conditioning,
163
+ old_eps=old_eps, t_next=ts_next,
164
+ dynamic_threshold=dynamic_threshold)
165
+ img, pred_x0, e_t = outs
166
+ old_eps.append(e_t)
167
+ if len(old_eps) >= 4:
168
+ old_eps.pop(0)
169
+ if callback: callback(i)
170
+ if img_callback: img_callback(pred_x0, i)
171
+
172
+ if index % log_every_t == 0 or index == total_steps - 1:
173
+ intermediates['x_inter'].append(img)
174
+ intermediates['pred_x0'].append(pred_x0)
175
+ return img, intermediates, None
176
+ def undo(self, x_t, t):
177
+ beta = extract_into_tensor(self.betas, t, x_t.shape)
178
+ x_t_forward = torch.sqrt(1 - beta) * x_t + torch.sqrt(beta) * torch.randn_like(x_t)
179
+ return x_t_forward
180
+
181
+ @torch.no_grad()
182
+ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
183
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
184
+ unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,
185
+ dynamic_threshold=None):
186
+ b, *_, device = *x.shape, x.device
187
+
188
+ def get_model_output(x, t):
189
+ if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
190
+ e_t, _ = self.model.apply_model(x, t, c)
191
+ else:
192
+ model_t, _ = self.model.apply_model(x,t,c)
193
+ model_uncond, _ = self.model.apply_model(x,t,unconditional_conditioning)
194
+
195
+ if isinstance(model_t, tuple):
196
+ model_t, _ = model_t
197
+ if isinstance(model_uncond, tuple):
198
+ model_uncond, _ = model_uncond
199
+ e_t = model_uncond + unconditional_guidance_scale * (model_t - model_uncond)
200
+
201
+ if score_corrector is not None:
202
+ assert self.model.parameterization == "eps"
203
+ e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
204
+
205
+ return e_t
206
+
207
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
208
+ alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
209
+ sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
210
+ sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
211
+
212
+ def get_x_prev_and_pred_x0(e_t, index):
213
+ # select parameters corresponding to the currently considered timestep
214
+ a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
215
+ a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
216
+ sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
217
+ sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
218
+
219
+ # current prediction for x_0
220
+ pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
221
+ if quantize_denoised:
222
+ pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
223
+ if dynamic_threshold is not None:
224
+ pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
225
+ # direction pointing to x_t
226
+ dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
227
+ noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
228
+ if noise_dropout > 0.:
229
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
230
+ x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
231
+ return x_prev, pred_x0
232
+
233
+ e_t = get_model_output(x, t)
234
+ if len(old_eps) == 0:
235
+ # Pseudo Improved Euler (2nd order)
236
+ x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
237
+ e_t_next = get_model_output(x_prev, t_next)
238
+ e_t_prime = (e_t + e_t_next) / 2
239
+ elif len(old_eps) == 1:
240
+ # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
241
+ e_t_prime = (3 * e_t - old_eps[-1]) / 2
242
+ elif len(old_eps) == 2:
243
+ # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
244
+ e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
245
+ elif len(old_eps) >= 3:
246
+ # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
247
+ e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
248
+
249
+ x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
250
+
251
+ return x_prev, pred_x0, e_t
cldm/warping_cldm_network.py ADDED
@@ -0,0 +1,357 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch as th
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+
6
+ from ldm.modules.diffusionmodules.util import (
7
+ conv_nd,
8
+ linear,
9
+ zero_module,
10
+ timestep_embedding
11
+ )
12
+
13
+ from einops import rearrange
14
+ from ldm.modules.attention import SpatialTransformer
15
+ from ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock
16
+ from ldm.util import exists
17
+
18
+ class StableVITON(UNetModel):
19
+ def __init__(
20
+ self,
21
+ dim_head_denorm=1,
22
+ *args,
23
+ **kwargs,
24
+ ):
25
+ super().__init__(*args, **kwargs)
26
+ warp_flow_blks = []
27
+ warp_zero_convs = []
28
+
29
+ self.encode_output_chs = [
30
+ 320,
31
+ 320,
32
+ 640,
33
+ 640,
34
+ 640,
35
+ 1280,
36
+ 1280,
37
+ 1280,
38
+ 1280
39
+ ]
40
+
41
+ self.encode_output_chs2 = [
42
+ 320,
43
+ 320,
44
+ 320,
45
+ 320,
46
+ 640,
47
+ 640,
48
+ 640,
49
+ 1280,
50
+ 1280
51
+ ]
52
+
53
+
54
+ for in_ch, cont_ch in zip(self.encode_output_chs, self.encode_output_chs2):
55
+ dim_head = in_ch // self.num_heads
56
+ dim_head = dim_head // dim_head_denorm
57
+ warp_flow_blks.append(SpatialTransformer(
58
+ in_channels=in_ch,
59
+ n_heads=self.num_heads,
60
+ d_head=dim_head,
61
+ depth=self.transformer_depth,
62
+ context_dim=cont_ch,
63
+ use_linear=self.use_linear_in_transformer,
64
+ use_checkpoint=self.use_checkpoint,
65
+ ))
66
+ warp_zero_convs.append(self.make_zero_conv(in_ch))
67
+ self.warp_flow_blks = nn.ModuleList(reversed(warp_flow_blks))
68
+ self.warp_zero_convs = nn.ModuleList(reversed(warp_zero_convs))
69
+ def make_zero_conv(self, channels):
70
+ return zero_module(conv_nd(2, channels, channels, 1, padding=0))
71
+ def forward(self, x, timesteps=None, context=None, control=None, only_mid_control=False, **kwargs):
72
+ hs = []
73
+
74
+ with torch.no_grad():
75
+ t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
76
+ emb = self.time_embed(t_emb)
77
+ h = x.type(self.dtype)
78
+ for module in self.input_blocks:
79
+ h = module(h, emb, context)
80
+ hs.append(h)
81
+ h = self.middle_block(h, emb, context)
82
+
83
+ if control is not None:
84
+ hint = control.pop()
85
+
86
+ for module in self.output_blocks[:3]:
87
+ control.pop()
88
+ h = torch.cat([h, hs.pop()], dim=1)
89
+ h = module(h, emb, context)
90
+
91
+ n_warp = len(self.encode_output_chs)
92
+ for i, (module, warp_blk, warp_zc) in enumerate(zip(self.output_blocks[3:n_warp+3], self.warp_flow_blks, self.warp_zero_convs)):
93
+ if control is None or (h.shape[-2] == 8 and h.shape[-1] == 6):
94
+ assert 0, f"shape is wrong : {h.shape}"
95
+ else:
96
+ hint = control.pop()
97
+ h = self.warp(h, hint, warp_blk, warp_zc)
98
+ h = torch.cat([h, hs.pop()], dim=1)
99
+ h = module(h, emb, context)
100
+ for module in self.output_blocks[n_warp+3:]:
101
+ if control is None:
102
+ h = torch.cat([h, hs.pop()], dim=1)
103
+ else:
104
+ h = torch.cat([h, hs.pop()], dim=1)
105
+ h = module(h, emb, context)
106
+ h = h.type(x.dtype)
107
+ return self.out(h)
108
+ def warp(self, x, hint, crossattn_layer, zero_conv, mask1=None, mask2=None):
109
+ hint = rearrange(hint, "b c h w -> b (h w) c").contiguous()
110
+ output = crossattn_layer(x, hint)
111
+ output = zero_conv(output)
112
+ return output + x
113
+ class NoZeroConvControlNet(nn.Module):
114
+ def __init__(
115
+ self,
116
+ image_size,
117
+ in_channels,
118
+ model_channels,
119
+ hint_channels,
120
+ num_res_blocks,
121
+ attention_resolutions,
122
+ dropout=0,
123
+ channel_mult=(1, 2, 4, 8),
124
+ conv_resample=True,
125
+ dims=2,
126
+ use_checkpoint=False,
127
+ use_fp16=False,
128
+ num_heads=-1,
129
+ num_head_channels=-1,
130
+ num_heads_upsample=-1,
131
+ use_scale_shift_norm=False,
132
+ resblock_updown=False,
133
+ use_new_attention_order=False,
134
+ use_spatial_transformer=False, # custom transformer support
135
+ transformer_depth=1, # custom transformer support
136
+ context_dim=None, # custom transformer support
137
+ n_embed=None,
138
+ legacy=True,
139
+ disable_self_attentions=None,
140
+ num_attention_blocks=None,
141
+ disable_middle_self_attn=False,
142
+ use_linear_in_transformer=False,
143
+ use_VAEdownsample=False,
144
+ cond_first_ch=8,
145
+ ):
146
+ super().__init__()
147
+ if use_spatial_transformer:
148
+ assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
149
+
150
+ if context_dim is not None:
151
+ assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
152
+ from omegaconf.listconfig import ListConfig
153
+ if type(context_dim) == ListConfig:
154
+ context_dim = list(context_dim)
155
+
156
+ if num_heads_upsample == -1:
157
+ num_heads_upsample = num_heads
158
+
159
+ if num_heads == -1:
160
+ assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
161
+
162
+ if num_head_channels == -1:
163
+ assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
164
+
165
+ self.dims = dims
166
+ self.image_size = image_size
167
+ self.in_channels = in_channels
168
+ self.model_channels = model_channels
169
+ if isinstance(num_res_blocks, int):
170
+ self.num_res_blocks = len(channel_mult) * [num_res_blocks]
171
+ else:
172
+ if len(num_res_blocks) != len(channel_mult):
173
+ raise ValueError("provide num_res_blocks either as an int (globally constant) or "
174
+ "as a list/tuple (per-level) with the same length as channel_mult")
175
+ self.num_res_blocks = num_res_blocks
176
+ if disable_self_attentions is not None:
177
+ # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
178
+ assert len(disable_self_attentions) == len(channel_mult)
179
+ if num_attention_blocks is not None:
180
+ assert len(num_attention_blocks) == len(self.num_res_blocks)
181
+ assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
182
+ print(f"Constructor of UNetModel received um_attention_blocks={num_attention_blocks}. "
183
+ f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
184
+ f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
185
+ f"attention will still not be set.")
186
+
187
+ self.attention_resolutions = attention_resolutions
188
+ self.dropout = dropout
189
+ self.channel_mult = channel_mult
190
+ self.conv_resample = conv_resample
191
+ self.use_checkpoint = use_checkpoint
192
+ self.dtype = th.float16 if use_fp16 else th.float32
193
+ self.num_heads = num_heads
194
+ self.num_head_channels = num_head_channels
195
+ self.num_heads_upsample = num_heads_upsample
196
+ self.predict_codebook_ids = n_embed is not None
197
+ self.use_VAEdownsample = use_VAEdownsample
198
+
199
+ time_embed_dim = model_channels * 4
200
+ self.time_embed = nn.Sequential(
201
+ linear(model_channels, time_embed_dim),
202
+ nn.SiLU(),
203
+ linear(time_embed_dim, time_embed_dim),
204
+ )
205
+
206
+ self.input_blocks = nn.ModuleList(
207
+ [
208
+ TimestepEmbedSequential(
209
+ conv_nd(dims, in_channels, model_channels, 3, padding=1)
210
+ )
211
+ ]
212
+ )
213
+
214
+ self.cond_first_block = TimestepEmbedSequential(
215
+ zero_module(conv_nd(dims, cond_first_ch, model_channels, 3, padding=1))
216
+ )
217
+
218
+
219
+ self._feature_size = model_channels
220
+ input_block_chans = [model_channels]
221
+ ch = model_channels
222
+ ds = 1
223
+ for level, mult in enumerate(channel_mult):
224
+ for nr in range(self.num_res_blocks[level]):
225
+ layers = [
226
+ ResBlock(
227
+ ch,
228
+ time_embed_dim,
229
+ dropout,
230
+ out_channels=mult * model_channels,
231
+ dims=dims,
232
+ use_checkpoint=use_checkpoint,
233
+ use_scale_shift_norm=use_scale_shift_norm,
234
+ )
235
+ ]
236
+ ch = mult * model_channels
237
+ if ds in attention_resolutions:
238
+ if num_head_channels == -1:
239
+ dim_head = ch // num_heads
240
+ else:
241
+ num_heads = ch // num_head_channels
242
+ dim_head = num_head_channels
243
+ if legacy:
244
+ # num_heads = 1
245
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
246
+ if exists(disable_self_attentions):
247
+ disabled_sa = disable_self_attentions[level]
248
+ else:
249
+ disabled_sa = False
250
+
251
+ if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
252
+ layers.append(
253
+ AttentionBlock(
254
+ ch,
255
+ use_checkpoint=use_checkpoint,
256
+ num_heads=num_heads,
257
+ num_head_channels=dim_head,
258
+ use_new_attention_order=use_new_attention_order,
259
+ ) if not use_spatial_transformer else SpatialTransformer(
260
+ ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
261
+ disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
262
+ use_checkpoint=use_checkpoint
263
+ )
264
+ )
265
+ self.input_blocks.append(TimestepEmbedSequential(*layers))
266
+ self._feature_size += ch
267
+ input_block_chans.append(ch)
268
+ if level != len(channel_mult) - 1:
269
+ out_ch = ch
270
+ self.input_blocks.append(
271
+ TimestepEmbedSequential(
272
+ ResBlock(
273
+ ch,
274
+ time_embed_dim,
275
+ dropout,
276
+ out_channels=out_ch,
277
+ dims=dims,
278
+ use_checkpoint=use_checkpoint,
279
+ use_scale_shift_norm=use_scale_shift_norm,
280
+ down=True,
281
+ )
282
+ if resblock_updown
283
+ else Downsample(
284
+ ch, conv_resample, dims=dims, out_channels=out_ch
285
+ )
286
+ )
287
+ )
288
+ ch = out_ch
289
+ input_block_chans.append(ch)
290
+ ds *= 2
291
+ self._feature_size += ch
292
+
293
+ if num_head_channels == -1:
294
+ dim_head = ch // num_heads
295
+ else:
296
+ num_heads = ch // num_head_channels
297
+ dim_head = num_head_channels
298
+ if legacy:
299
+ # num_heads = 1
300
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
301
+ self.middle_block = TimestepEmbedSequential(
302
+ ResBlock(
303
+ ch,
304
+ time_embed_dim,
305
+ dropout,
306
+ dims=dims,
307
+ use_checkpoint=use_checkpoint,
308
+ use_scale_shift_norm=use_scale_shift_norm,
309
+ ),
310
+ AttentionBlock(
311
+ ch,
312
+ use_checkpoint=use_checkpoint,
313
+ num_heads=num_heads,
314
+ num_head_channels=dim_head,
315
+ use_new_attention_order=use_new_attention_order,
316
+ ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn
317
+ ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
318
+ disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
319
+ use_checkpoint=use_checkpoint
320
+ ),
321
+ ResBlock(
322
+ ch,
323
+ time_embed_dim,
324
+ dropout,
325
+ dims=dims,
326
+ use_checkpoint=use_checkpoint,
327
+ use_scale_shift_norm=use_scale_shift_norm,
328
+ ),
329
+ )
330
+ self._feature_size += ch
331
+
332
+ def forward(self, x, hint, timesteps, context, only_mid_control=False, **kwargs):
333
+ t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
334
+ emb = self.time_embed(t_emb)
335
+
336
+ if not self.use_VAEdownsample:
337
+ guided_hint = self.input_hint_block(hint, emb, context)
338
+ else:
339
+ guided_hint = self.cond_first_block(hint, emb, context)
340
+
341
+ outs = []
342
+ hs = []
343
+ h = x.type(self.dtype)
344
+ for module in self.input_blocks:
345
+ if guided_hint is not None:
346
+ h = module(h, emb, context)
347
+ h += guided_hint
348
+ hs.append(h)
349
+ guided_hint = None
350
+ else:
351
+ h = module(h, emb, context)
352
+ hs.append(h)
353
+ outs.append(h)
354
+
355
+ h = self.middle_block(h, emb, context)
356
+ outs.append(h)
357
+ return outs, None
configs/VITON512.yaml ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: cldm.cldm.ControlLDM
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.0120
6
+ num_timesteps_cond: 1
7
+ log_every_t: 200
8
+ timesteps: 1000
9
+ first_stage_key: "image"
10
+ first_stage_key_cond: ["agn", "agn_mask", "image_densepose"]
11
+ cond_stage_key: "cloth"
12
+ control_key: "cloth"
13
+ image_size: 64
14
+ channels: 4
15
+ cond_stage_trainable: False
16
+ conditioning_key: crossattn
17
+ monitor: val/loss_simple_ema
18
+ scale_factor: 0.18215
19
+ use_ema: False
20
+ only_mid_control: False
21
+ use_VAEdownsample: True
22
+ use_lastzc: True
23
+ use_imageCLIP: True
24
+ use_pbe_weight: True
25
+ u_cond_percent: 0.2
26
+ use_attn_mask: False
27
+ mask1_key: "agn_mask"
28
+ mask2_key: "cloth_mask"
29
+
30
+ control_stage_config:
31
+ target: cldm.warping_cldm_network.NoZeroConvControlNet
32
+ params:
33
+ image_size: 32
34
+ in_channels: 13
35
+ hint_channels: 3
36
+ model_channels: 320
37
+ attention_resolutions: [ 4, 2, 1 ]
38
+ num_res_blocks: 2
39
+ channel_mult: [ 1, 2, 4, 4 ]
40
+ num_heads: 8
41
+ use_spatial_transformer: True
42
+ transformer_depth: 1
43
+ context_dim: 768
44
+ use_checkpoint: True
45
+ legacy: False
46
+ cond_first_ch: 4
47
+
48
+ unet_config:
49
+ target: cldm.warping_cldm_network.StableVITON
50
+ params:
51
+ image_size: 32
52
+ in_channels: 13
53
+ out_channels: 4
54
+ model_channels: 320
55
+ attention_resolutions: [ 4, 2, 1 ]
56
+ num_res_blocks: 2
57
+ channel_mult: [ 1, 2, 4, 4 ]
58
+ num_heads: 8
59
+ use_spatial_transformer: True
60
+ transformer_depth: 1
61
+ context_dim: 768
62
+ use_checkpoint: True
63
+ legacy: False
64
+ dim_head_denorm: 1
65
+
66
+ first_stage_config:
67
+ target: ldm.models.autoencoder.AutoencoderKL
68
+ params:
69
+ embed_dim: 4
70
+ monitor: val/rec_loss
71
+ ddconfig:
72
+ double_z: true
73
+ z_channels: 4
74
+ resolution: 256
75
+ in_channels: 3
76
+ out_ch: 3
77
+ ch: 128
78
+ ch_mult:
79
+ - 1
80
+ - 2
81
+ - 4
82
+ - 4
83
+ num_res_blocks: 2
84
+ attn_resolutions: []
85
+ dropout: 0.0
86
+ lossconfig:
87
+ target: torch.nn.Identity
88
+ validation_config:
89
+ ddim_steps: 50
90
+ eta: 0.0
91
+ scale: 1.0
92
+
93
+ cond_stage_config:
94
+ target: ldm.modules.image_encoders.modules.FrozenCLIPImageEmbedder
95
+ dataset_name: VITONHDDataset
96
+ resume_path: ./pretrained_models/VITONHD_PBE_pose.ckpt
97
+ default_prompt: ""
98
+ log_images_kwargs:
99
+ unconditional_guidance_scale: 5.0
100
+
ldm/data/__init__.py ADDED
File without changes
ldm/data/util.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ from ldm.modules.midas.api import load_midas_transform
4
+
5
+
6
+ class AddMiDaS(object):
7
+ def __init__(self, model_type):
8
+ super().__init__()
9
+ self.transform = load_midas_transform(model_type)
10
+
11
+ def pt2np(self, x):
12
+ x = ((x + 1.0) * .5).detach().cpu().numpy()
13
+ return x
14
+
15
+ def np2pt(self, x):
16
+ x = torch.from_numpy(x) * 2 - 1.
17
+ return x
18
+
19
+ def __call__(self, sample):
20
+ # sample['jpg'] is tensor hwc in [-1, 1] at this point
21
+ x = self.pt2np(sample['jpg'])
22
+ x = self.transform({"image": x})["image"]
23
+ sample['midas_in'] = x
24
+ return sample
ldm/models/autoencoder.py ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import pytorch_lightning as pl
3
+ import torch.nn.functional as F
4
+ from contextlib import contextmanager
5
+
6
+ from ldm.modules.diffusionmodules.model import Encoder, Decoder
7
+ from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
8
+
9
+ from ldm.util import instantiate_from_config
10
+ from ldm.modules.ema import LitEma
11
+
12
+
13
+ class AutoencoderKL(pl.LightningModule):
14
+ def __init__(self,
15
+ ddconfig,
16
+ lossconfig,
17
+ embed_dim,
18
+ ckpt_path=None,
19
+ ignore_keys=[],
20
+ image_key="image",
21
+ colorize_nlabels=None,
22
+ monitor=None,
23
+ ema_decay=None,
24
+ learn_logvar=False
25
+ ):
26
+ super().__init__()
27
+ self.lossconfig = lossconfig
28
+ self.learn_logvar = learn_logvar
29
+ self.image_key = image_key
30
+ self.encoder = Encoder(**ddconfig)
31
+ self.decoder = Decoder(**ddconfig)
32
+ self.loss = torch.nn.Identity()
33
+ assert ddconfig["double_z"]
34
+ self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
35
+ self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
36
+ self.embed_dim = embed_dim
37
+ if colorize_nlabels is not None:
38
+ assert type(colorize_nlabels)==int
39
+ self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
40
+ if monitor is not None:
41
+ self.monitor = monitor
42
+
43
+ self.use_ema = ema_decay is not None
44
+ if self.use_ema:
45
+ self.ema_decay = ema_decay
46
+ assert 0. < ema_decay < 1.
47
+ self.model_ema = LitEma(self, decay=ema_decay)
48
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
49
+
50
+ if ckpt_path is not None:
51
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
52
+ def init_loss(self):
53
+ self.loss = instantiate_from_config(self.lossconfig)
54
+ def init_from_ckpt(self, path, ignore_keys=list()):
55
+ sd = torch.load(path, map_location="cpu")["state_dict"]
56
+ keys = list(sd.keys())
57
+ for k in keys:
58
+ for ik in ignore_keys:
59
+ if k.startswith(ik):
60
+ print("Deleting key {} from state_dict.".format(k))
61
+ del sd[k]
62
+ self.load_state_dict(sd, strict=False)
63
+ print(f"Restored from {path}")
64
+
65
+ @contextmanager
66
+ def ema_scope(self, context=None):
67
+ if self.use_ema:
68
+ self.model_ema.store(self.parameters())
69
+ self.model_ema.copy_to(self)
70
+ if context is not None:
71
+ print(f"{context}: Switched to EMA weights")
72
+ try:
73
+ yield None
74
+ finally:
75
+ if self.use_ema:
76
+ self.model_ema.restore(self.parameters())
77
+ if context is not None:
78
+ print(f"{context}: Restored training weights")
79
+
80
+ def on_train_batch_end(self, *args, **kwargs):
81
+ if self.use_ema:
82
+ self.model_ema(self)
83
+
84
+ def encode(self, x):
85
+ h = self.encoder(x)
86
+ moments = self.quant_conv(h)
87
+ posterior = DiagonalGaussianDistribution(moments)
88
+ return posterior
89
+
90
+ def decode(self, z):
91
+ z = self.post_quant_conv(z)
92
+ dec = self.decoder(z)
93
+ return dec
94
+
95
+ def forward(self, input, sample_posterior=True):
96
+ posterior = self.encode(input)
97
+ if sample_posterior:
98
+ z = posterior.sample()
99
+ else:
100
+ z = posterior.mode()
101
+ dec = self.decode(z)
102
+ return dec, posterior
103
+
104
+ def get_input(self, batch, k):
105
+ x = batch[k]
106
+ if len(x.shape) == 3:
107
+ x = x[..., None]
108
+ x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
109
+ return x
110
+
111
+ def training_step(self, batch, batch_idx):
112
+ real_img = self.get_input(batch, self.image_key)
113
+ recon, posterior = self(real_img)
114
+ loss = self.loss(real_img, recon, posterior)
115
+ return loss
116
+
117
+ def validation_step(self, batch, batch_idx):
118
+ log_dict = self._validation_step(batch, batch_idx)
119
+ with self.ema_scope():
120
+ log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema")
121
+ return log_dict
122
+
123
+ def _validation_step(self, batch, batch_idx, postfix=""):
124
+ inputs = self.get_input(batch, self.image_key)
125
+ reconstructions, posterior = self(inputs)
126
+ aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
127
+ last_layer=self.get_last_layer(), split="val"+postfix)
128
+
129
+ discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
130
+ last_layer=self.get_last_layer(), split="val"+postfix)
131
+
132
+ self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"])
133
+ self.log_dict(log_dict_ae)
134
+ self.log_dict(log_dict_disc)
135
+ return self.log_dict
136
+ def configure_optimizers(self):
137
+ lr = self.learning_rate
138
+ ae_params_list = list(self.decoder.parameters())
139
+ if self.learn_logvar:
140
+ print(f"{self.__class__.__name__}: Learning logvar")
141
+ ae_params_list.append(self.loss.logvar)
142
+ opt_ae = torch.optim.Adam(ae_params_list,
143
+ lr=lr, betas=(0.5, 0.9))
144
+ return [opt_ae], []
145
+
146
+ def get_last_layer(self):
147
+ return self.decoder.conv_out.weight
148
+
149
+ @torch.no_grad()
150
+ def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs):
151
+ log = dict()
152
+ x = self.get_input(batch, self.image_key)
153
+ x = x.to(self.device)
154
+ if not only_inputs:
155
+ xrec, posterior = self(x)
156
+ if x.shape[1] > 3:
157
+ # colorize with random projection
158
+ assert xrec.shape[1] > 3
159
+ x = self.to_rgb(x)
160
+ xrec = self.to_rgb(xrec)
161
+ log["samples"] = self.decode(torch.randn_like(posterior.sample()))
162
+ log["reconstructions"] = xrec
163
+ if log_ema or self.use_ema:
164
+ with self.ema_scope():
165
+ xrec_ema, posterior_ema = self(x)
166
+ if x.shape[1] > 3:
167
+ # colorize with random projection
168
+ assert xrec_ema.shape[1] > 3
169
+ xrec_ema = self.to_rgb(xrec_ema)
170
+ log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample()))
171
+ log["reconstructions_ema"] = xrec_ema
172
+ log["inputs"] = x
173
+ return log
174
+
175
+ def to_rgb(self, x):
176
+ assert self.image_key == "segmentation"
177
+ if not hasattr(self, "colorize"):
178
+ self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
179
+ x = F.conv2d(x, weight=self.colorize)
180
+ x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
181
+ return x
182
+
183
+
184
+ class IdentityFirstStage(torch.nn.Module):
185
+ def __init__(self, *args, vq_interface=False, **kwargs):
186
+ self.vq_interface = vq_interface
187
+ super().__init__()
188
+
189
+ def encode(self, x, *args, **kwargs):
190
+ return x
191
+
192
+ def decode(self, x, *args, **kwargs):
193
+ return x
194
+
195
+ def quantize(self, x, *args, **kwargs):
196
+ if self.vq_interface:
197
+ return x, None, [None, None, None]
198
+ return x
199
+
200
+ def forward(self, x, *args, **kwargs):
201
+ return x
202
+
ldm/models/diffusion/__init__.py ADDED
File without changes
ldm/models/diffusion/ddim.py ADDED
@@ -0,0 +1,377 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """SAMPLING ONLY."""
2
+
3
+ import torch
4
+ import numpy as np
5
+ from tqdm import tqdm
6
+
7
+ from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor
8
+
9
+
10
+ class DDIMSampler(object):
11
+ def __init__(self, model, schedule="linear", **kwargs):
12
+ super().__init__()
13
+ self.model = model
14
+ self.ddpm_num_timesteps = model.num_timesteps
15
+ self.schedule = schedule
16
+
17
+ def register_buffer(self, name, attr):
18
+ if type(attr) == torch.Tensor:
19
+ if attr.device != torch.device("cuda"):
20
+ attr = attr.to(torch.device("cuda"))
21
+ setattr(self, name, attr)
22
+
23
+ def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
24
+ self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
25
+ num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
26
+ alphas_cumprod = self.model.alphas_cumprod
27
+ assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
28
+ to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
29
+
30
+ self.register_buffer('betas', to_torch(self.model.betas))
31
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
32
+ self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
33
+
34
+ # calculations for diffusion q(x_t | x_{t-1}) and others
35
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
36
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
37
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
38
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
39
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
40
+
41
+ # ddim sampling parameters
42
+ ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
43
+ ddim_timesteps=self.ddim_timesteps,
44
+ eta=ddim_eta,verbose=verbose)
45
+ self.register_buffer('ddim_sigmas', ddim_sigmas)
46
+ self.register_buffer('ddim_alphas', ddim_alphas)
47
+ self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
48
+ self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
49
+ sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
50
+ (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
51
+ 1 - self.alphas_cumprod / self.alphas_cumprod_prev))
52
+ self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
53
+
54
+ @torch.no_grad()
55
+ def sample(self,
56
+ S,
57
+ batch_size,
58
+ shape,
59
+ conditioning=None,
60
+ callback=None,
61
+ normals_sequence=None,
62
+ img_callback=None,
63
+ quantize_x0=False,
64
+ eta=0.,
65
+ mask=None,
66
+ x0=None,
67
+ temperature=1.,
68
+ noise_dropout=0.,
69
+ score_corrector=None,
70
+ corrector_kwargs=None,
71
+ verbose=True,
72
+ x_T=None,
73
+ log_every_t=100,
74
+ unconditional_guidance_scale=1.,
75
+ unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
76
+ dynamic_threshold=None,
77
+ ucg_schedule=None,
78
+ **kwargs
79
+ ):
80
+ if conditioning is not None:
81
+ if isinstance(conditioning, dict):
82
+ ctmp = conditioning[list(conditioning.keys())[0]]
83
+ while isinstance(ctmp, list): ctmp = ctmp[0]
84
+ cbs = ctmp.shape[0]
85
+ if cbs != batch_size:
86
+ print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
87
+
88
+ elif isinstance(conditioning, list):
89
+ for ctmp in conditioning:
90
+ if ctmp.shape[0] != batch_size:
91
+ print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
92
+
93
+ else:
94
+ if conditioning.shape[0] != batch_size:
95
+ print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
96
+
97
+ self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
98
+ # sampling
99
+ C, H, W = shape
100
+ size = (batch_size, C, H, W)
101
+ print(f'Data shape for DDIM sampling is {size}, eta {eta}')
102
+
103
+ samples, intermediates, cond_output_dict = self.ddim_sampling(conditioning, size,
104
+ callback=callback,
105
+ img_callback=img_callback,
106
+ quantize_denoised=quantize_x0,
107
+ mask=mask, x0=x0,
108
+ ddim_use_original_steps=False,
109
+ noise_dropout=noise_dropout,
110
+ temperature=temperature,
111
+ score_corrector=score_corrector,
112
+ corrector_kwargs=corrector_kwargs,
113
+ x_T=x_T,
114
+ log_every_t=log_every_t,
115
+ unconditional_guidance_scale=unconditional_guidance_scale,
116
+ unconditional_conditioning=unconditional_conditioning,
117
+ dynamic_threshold=dynamic_threshold,
118
+ ucg_schedule=ucg_schedule
119
+ )
120
+ return samples, intermediates, cond_output_dict
121
+
122
+ @torch.no_grad()
123
+ def ddim_sampling(self, cond, shape,
124
+ x_T=None, ddim_use_original_steps=False,
125
+ callback=None, timesteps=None, quantize_denoised=False,
126
+ mask=None, x0=None, img_callback=None, log_every_t=100,
127
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
128
+ unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None,
129
+ ucg_schedule=None):
130
+ device = self.model.betas.device
131
+ b = shape[0]
132
+ if x_T is None:
133
+ img = torch.randn(shape, device=device)
134
+ else:
135
+ img = x_T
136
+
137
+ if timesteps is None:
138
+ timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
139
+ elif timesteps is not None and not ddim_use_original_steps:
140
+ subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
141
+ timesteps = self.ddim_timesteps[:subset_end]
142
+
143
+ intermediates = {'x_inter': [img], 'pred_x0': [img]}
144
+ time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
145
+ total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
146
+ print(f"Running DDIM Sampling with {total_steps} timesteps")
147
+
148
+ iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
149
+
150
+ for i, step in enumerate(iterator):
151
+ index = total_steps - i - 1
152
+ ts = torch.full((b,), step, device=device, dtype=torch.long)
153
+
154
+ if mask is not None:
155
+ assert x0 is not None
156
+ img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
157
+ img = img_orig * mask + (1. - mask) * img
158
+
159
+ if ucg_schedule is not None:
160
+ assert len(ucg_schedule) == len(time_range)
161
+ unconditional_guidance_scale = ucg_schedule[i]
162
+
163
+ outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
164
+ quantize_denoised=quantize_denoised, temperature=temperature,
165
+ noise_dropout=noise_dropout, score_corrector=score_corrector,
166
+ corrector_kwargs=corrector_kwargs,
167
+ unconditional_guidance_scale=unconditional_guidance_scale,
168
+ unconditional_conditioning=unconditional_conditioning,
169
+ dynamic_threshold=dynamic_threshold)
170
+ img, pred_x0, cond_output_dict = outs
171
+ if callback: callback(i)
172
+ if img_callback: img_callback(pred_x0, i)
173
+
174
+ if index % log_every_t == 0 or index == total_steps - 1:
175
+ intermediates['x_inter'].append(img)
176
+ intermediates['pred_x0'].append(pred_x0)
177
+
178
+ if cond_output_dict is not None:
179
+ cond_output = cond_output_dict["cond_output"]
180
+ if self.model.use_noisy_cond:
181
+ b = cond_output.shape[0]
182
+
183
+ alphas = self.model.alphas_cumprod if ddim_use_original_steps else self.ddim_alphas
184
+ alphas_prev = self.model.alphas_cumprod_prev if ddim_use_original_steps else self.ddim_alphas_prev
185
+ sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if ddim_use_original_steps else self.ddim_sqrt_one_minus_alphas
186
+ sigmas = self.model.ddim_sigmas_for_original_num_steps if ddim_use_original_steps else self.ddim_sigmas
187
+
188
+ device = cond_output.device
189
+ a_t = torch.full((b, 1, 1, 1), alphas[0], device=device)
190
+ a_prev = torch.full((b, 1, 1, 1), alphas_prev[0], device=device)
191
+ sigma_t = torch.full((b, 1, 1, 1), sigmas[0], device=device)
192
+ sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[0], device=device)
193
+
194
+ c = cond_output_dict["cond_input"]
195
+ e_t = cond_output
196
+ pred_c0 = (c - sqrt_one_minus_at * e_t) / a_t.sqrt()
197
+ dir_ct = (1. - a_prev - sigma_t**2).sqrt() * e_t
198
+ noise = sigma_t * noise_like(c.shape, device, False) * temperature
199
+
200
+ if noise_dropout > 0.:
201
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
202
+ cond_output = a_prev.sqrt() * pred_c0 + dir_ct + noise
203
+ cond_output_dict[f"cond_sample"] = cond_output
204
+ return img, intermediates, cond_output_dict
205
+
206
+ @torch.no_grad()
207
+ def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
208
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
209
+ unconditional_guidance_scale=1., unconditional_conditioning=None,
210
+ dynamic_threshold=None):
211
+ b, *_, device = *x.shape, x.device
212
+ if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
213
+ model_output, cond_output_dict = self.model.apply_model(x, t, c)
214
+ else:
215
+ # x_in = torch.cat([x] * 2)
216
+ # t_in = torch.cat([t] * 2)
217
+ # if isinstance(c, dict):
218
+ # assert isinstance(unconditional_conditioning, dict)
219
+ # c_in = dict()
220
+ # for k in c:
221
+ # if isinstance(c[k], list):
222
+ # c_in[k] = [torch.cat([
223
+ # unconditional_conditioning[k][i],
224
+ # c[k][i]]) for i in range(len(c[k]))]
225
+ # else:
226
+ # c_in[k] = torch.cat([
227
+ # unconditional_conditioning[k],
228
+ # c[k]])
229
+ # elif isinstance(c, list):
230
+ # c_in = list()
231
+ # assert isinstance(unconditional_conditioning, list)
232
+ # for i in range(len(c)):
233
+ # c_in.append(torch.cat([unconditional_conditioning[i], c[i]]))
234
+ # else:
235
+ # c_in = torch.cat([unconditional_conditioning, c])
236
+ x_in = x
237
+ t_in = t
238
+ model_t, cond_output_dict_cond = self.model.apply_model(x_in, t_in, c)
239
+ model_uncond, cond_output_dict_uncond = self.model.apply_model(x_in, t_in, unconditional_conditioning)
240
+ if isinstance(model_t, tuple):
241
+ model_t, _ = model_t
242
+ if isinstance(model_uncond, tuple):
243
+ model_uncond, _ = model_uncond
244
+ if cond_output_dict_cond is not None:
245
+ cond_output_dict = dict()
246
+ for k in cond_output_dict_cond.keys():
247
+ cond_output_dict[k] = torch.cat([cond_output_dict_uncond[k], cond_output_dict_cond[k]])
248
+ else:
249
+ cond_output_dict = None
250
+ # model_output, cond_output_dict = self.model.apply_model(x_in, t_in, c_in)
251
+ # model_uncond, model_t = model_output.chunk(2)
252
+ model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond)
253
+
254
+ if self.model.parameterization == "v":
255
+ e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
256
+ else:
257
+ e_t = model_output
258
+
259
+ if score_corrector is not None:
260
+ assert self.model.parameterization == "eps", 'not implemented'
261
+ e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
262
+
263
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
264
+ alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
265
+ sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
266
+ sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
267
+ # select parameters corresponding to the currently considered timestep
268
+ a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
269
+ a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
270
+ sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
271
+ sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
272
+
273
+ # current prediction for x_0
274
+ if self.model.parameterization != "v":
275
+ pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
276
+ else:
277
+ pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
278
+
279
+ if quantize_denoised:
280
+ pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
281
+
282
+ if dynamic_threshold is not None:
283
+ raise NotImplementedError()
284
+
285
+ # direction pointing to x_t
286
+ dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
287
+ noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
288
+ if noise_dropout > 0.:
289
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
290
+ x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
291
+
292
+ return x_prev, pred_x0, cond_output_dict
293
+
294
+ @torch.no_grad()
295
+ def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None,
296
+ unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None):
297
+ num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0]
298
+
299
+ assert t_enc <= num_reference_steps
300
+ num_steps = t_enc
301
+
302
+ if use_original_steps:
303
+ alphas_next = self.alphas_cumprod[:num_steps]
304
+ alphas = self.alphas_cumprod_prev[:num_steps]
305
+ else:
306
+ alphas_next = self.ddim_alphas[:num_steps]
307
+ alphas = torch.tensor(self.ddim_alphas_prev[:num_steps])
308
+
309
+ x_next = x0
310
+ intermediates = []
311
+ inter_steps = []
312
+ for i in tqdm(range(num_steps), desc='Encoding Image'):
313
+ t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long)
314
+ if unconditional_guidance_scale == 1.:
315
+ noise_pred = self.model.apply_model(x_next, t, c)[0]
316
+ else:
317
+ assert unconditional_conditioning is not None
318
+ e_t_uncond, noise_pred = torch.chunk(
319
+ self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)),
320
+ torch.cat((unconditional_conditioning, c))), 2)
321
+ noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond)[0]
322
+
323
+ xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next
324
+ weighted_noise_pred = alphas_next[i].sqrt() * (
325
+ (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred
326
+ x_next = xt_weighted + weighted_noise_pred
327
+ if return_intermediates and i % (
328
+ num_steps // return_intermediates) == 0 and i < num_steps - 1:
329
+ intermediates.append(x_next)
330
+ inter_steps.append(i)
331
+ elif return_intermediates and i >= num_steps - 2:
332
+ intermediates.append(x_next)
333
+ inter_steps.append(i)
334
+ if callback: callback(i)
335
+
336
+ out = {'x_encoded': x_next, 'intermediate_steps': inter_steps}
337
+ if return_intermediates:
338
+ out.update({'intermediates': intermediates})
339
+ return x_next, out
340
+
341
+ @torch.no_grad()
342
+ def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
343
+ # fast, but does not allow for exact reconstruction
344
+ # t serves as an index to gather the correct alphas
345
+ if use_original_steps:
346
+ sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
347
+ sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
348
+ else:
349
+ sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
350
+ sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
351
+
352
+ if noise is None:
353
+ noise = torch.randn_like(x0)
354
+ return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
355
+ extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
356
+
357
+ @torch.no_grad()
358
+ def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
359
+ use_original_steps=False, callback=None):
360
+
361
+ timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
362
+ timesteps = timesteps[:t_start]
363
+
364
+ time_range = np.flip(timesteps)
365
+ total_steps = timesteps.shape[0]
366
+ print(f"Running DDIM Sampling with {total_steps} timesteps")
367
+
368
+ iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
369
+ x_dec = x_latent
370
+ for i, step in enumerate(iterator):
371
+ index = total_steps - i - 1
372
+ ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
373
+ x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
374
+ unconditional_guidance_scale=unconditional_guidance_scale,
375
+ unconditional_conditioning=unconditional_conditioning)
376
+ if callback: callback(i)
377
+ return x_dec
ldm/models/diffusion/ddpm.py ADDED
@@ -0,0 +1,1875 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ wild mixture of
3
+ https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
4
+ https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
5
+ https://github.com/CompVis/taming-transformers
6
+ -- merci
7
+ """
8
+
9
+ import torch
10
+ import torch.nn as nn
11
+ import numpy as np
12
+ import pytorch_lightning as pl
13
+ from torch.optim.lr_scheduler import LambdaLR
14
+ from einops import rearrange, repeat
15
+ from contextlib import contextmanager, nullcontext
16
+ from functools import partial
17
+ import itertools
18
+ from tqdm import tqdm
19
+ from torchvision.utils import make_grid
20
+ from pytorch_lightning.utilities.distributed import rank_zero_only
21
+ from omegaconf import ListConfig
22
+ from torchvision.transforms.functional import resize
23
+ import torchvision.transforms as T
24
+ import random
25
+ import torch.nn.functional as F
26
+ from diffusers.models.autoencoder_kl import AutoencoderKLOutput
27
+ from diffusers.models.vae import DecoderOutput
28
+
29
+ from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
30
+ from ldm.modules.ema import LitEma
31
+ from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
32
+ from ldm.models.autoencoder import IdentityFirstStage, AutoencoderKL
33
+ from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like, zero_module, conv_nd
34
+ from ldm.models.diffusion.ddim import DDIMSampler
35
+
36
+ __conditioning_keys__ = {'concat': 'c_concat',
37
+ 'crossattn': 'c_crossattn',
38
+ 'adm': 'y'}
39
+
40
+
41
+ def disabled_train(self, mode=True):
42
+ """Overwrite model.train with this function to make sure train/eval mode
43
+ does not change anymore."""
44
+ return self
45
+
46
+
47
+ def uniform_on_device(r1, r2, shape, device):
48
+ return (r1 - r2) * torch.rand(*shape, device=device) + r2
49
+
50
+ class DDPM(pl.LightningModule):
51
+ # classic DDPM with Gaussian diffusion, in image space
52
+ def __init__(self,
53
+ unet_config,
54
+ timesteps=1000,
55
+ beta_schedule="linear",
56
+ loss_type="l2",
57
+ ckpt_path=None,
58
+ ignore_keys=[],
59
+ load_only_unet=False,
60
+ monitor="val/loss",
61
+ use_ema=True,
62
+ first_stage_key="image",
63
+ image_size=256,
64
+ channels=3,
65
+ log_every_t=100,
66
+ clip_denoised=True,
67
+ linear_start=1e-4,
68
+ linear_end=2e-2,
69
+ cosine_s=8e-3,
70
+ given_betas=None,
71
+ original_elbo_weight=0.,
72
+ v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
73
+ l_simple_weight=1.,
74
+ conditioning_key=None,
75
+ parameterization="eps", # all assuming fixed variance schedules
76
+ scheduler_config=None,
77
+ use_positional_encodings=False,
78
+ learn_logvar=False,
79
+ logvar_init=0.,
80
+ make_it_fit=False,
81
+ ucg_training=None,
82
+ reset_ema=False,
83
+ reset_num_ema_updates=False,
84
+ l_cond_simple_weight=1.0,
85
+ l_cond_recon_weight=1.0,
86
+ **kwargs
87
+ ):
88
+ super().__init__()
89
+ assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"'
90
+ self.parameterization = parameterization
91
+ print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
92
+ self.unet_config = unet_config
93
+ self.cond_stage_model = None
94
+ self.clip_denoised = clip_denoised
95
+ self.log_every_t = log_every_t
96
+ self.first_stage_key = first_stage_key
97
+ self.image_size = image_size # try conv?
98
+ self.channels = channels
99
+ self.use_positional_encodings = use_positional_encodings
100
+ self.model = DiffusionWrapper(unet_config, conditioning_key)
101
+ count_params(self.model, verbose=True)
102
+ self.use_ema = use_ema
103
+ if self.use_ema:
104
+ self.model_ema = LitEma(self.model)
105
+ print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
106
+
107
+ self.use_scheduler = scheduler_config is not None
108
+ if self.use_scheduler:
109
+ self.scheduler_config = scheduler_config
110
+ self.imagenet_norm = T.Normalize((0.48145466, 0.4578275, 0.40821073),
111
+ (0.26862954, 0.26130258, 0.27577711))
112
+
113
+ self.v_posterior = v_posterior
114
+ self.original_elbo_weight = original_elbo_weight
115
+ self.l_simple_weight = l_simple_weight
116
+ self.l_cond_simple_weight = l_cond_simple_weight
117
+ self.l_cond_recon_weight = l_cond_recon_weight
118
+
119
+ if monitor is not None:
120
+ self.monitor = monitor
121
+ self.make_it_fit = make_it_fit
122
+ if reset_ema: assert exists(ckpt_path)
123
+ if ckpt_path is not None:
124
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
125
+ if reset_ema:
126
+ assert self.use_ema
127
+ print(f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.")
128
+ self.model_ema = LitEma(self.model)
129
+ if reset_num_ema_updates:
130
+ print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ")
131
+ assert self.use_ema
132
+ self.model_ema.reset_num_updates()
133
+
134
+ self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
135
+ linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
136
+
137
+ self.loss_type = loss_type
138
+
139
+ self.learn_logvar = learn_logvar
140
+ logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
141
+ if self.learn_logvar:
142
+ self.logvar = nn.Parameter(self.logvar, requires_grad=True)
143
+ else:
144
+ self.register_buffer('logvar', logvar)
145
+
146
+ self.ucg_training = ucg_training or dict()
147
+ if self.ucg_training:
148
+ self.ucg_prng = np.random.RandomState()
149
+
150
+ def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
151
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
152
+ if exists(given_betas):
153
+ betas = given_betas
154
+ else:
155
+ betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
156
+ cosine_s=cosine_s)
157
+ alphas = 1. - betas
158
+ alphas_cumprod = np.cumprod(alphas, axis=0)
159
+ alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
160
+
161
+ timesteps, = betas.shape
162
+ self.num_timesteps = int(timesteps)
163
+ self.linear_start = linear_start
164
+ self.linear_end = linear_end
165
+ assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
166
+
167
+ to_torch = partial(torch.tensor, dtype=torch.float32)
168
+
169
+ self.register_buffer('betas', to_torch(betas))
170
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
171
+ self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
172
+
173
+ # calculations for diffusion q(x_t | x_{t-1}) and others
174
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
175
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
176
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
177
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
178
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
179
+
180
+ # calculations for posterior q(x_{t-1} | x_t, x_0)
181
+ posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
182
+ 1. - alphas_cumprod) + self.v_posterior * betas
183
+ # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
184
+ self.register_buffer('posterior_variance', to_torch(posterior_variance))
185
+ # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
186
+ self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
187
+ self.register_buffer('posterior_mean_coef1', to_torch(
188
+ betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
189
+ self.register_buffer('posterior_mean_coef2', to_torch(
190
+ (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
191
+
192
+ if self.parameterization == "eps":
193
+ lvlb_weights = self.betas ** 2 / (
194
+ 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
195
+ elif self.parameterization == "x0":
196
+ lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
197
+ elif self.parameterization == "v":
198
+ lvlb_weights = torch.ones_like(self.betas ** 2 / (
199
+ 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)))
200
+ else:
201
+ raise NotImplementedError("mu not supported")
202
+ lvlb_weights[0] = lvlb_weights[1]
203
+ self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
204
+ assert not torch.isnan(self.lvlb_weights).all()
205
+
206
+ @contextmanager
207
+ def ema_scope(self, context=None):
208
+ if self.use_ema:
209
+ self.model_ema.store(self.model.parameters())
210
+ self.model_ema.copy_to(self.model)
211
+ if context is not None:
212
+ print(f"{context}: Switched to EMA weights")
213
+ try:
214
+ yield None
215
+ finally:
216
+ if self.use_ema:
217
+ self.model_ema.restore(self.model.parameters())
218
+ if context is not None:
219
+ print(f"{context}: Restored training weights")
220
+
221
+ @torch.no_grad()
222
+ def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
223
+ sd = torch.load(path, map_location="cpu")
224
+ if "state_dict" in list(sd.keys()):
225
+ sd = sd["state_dict"]
226
+ keys = list(sd.keys())
227
+ for k in keys:
228
+ for ik in ignore_keys:
229
+ if k.startswith(ik):
230
+ print("Deleting key {} from state_dict.".format(k))
231
+ del sd[k]
232
+ if self.make_it_fit:
233
+ n_params = len([name for name, _ in
234
+ itertools.chain(self.named_parameters(),
235
+ self.named_buffers())])
236
+ for name, param in tqdm(
237
+ itertools.chain(self.named_parameters(),
238
+ self.named_buffers()),
239
+ desc="Fitting old weights to new weights",
240
+ total=n_params
241
+ ):
242
+ if not name in sd:
243
+ continue
244
+ old_shape = sd[name].shape
245
+ new_shape = param.shape
246
+ assert len(old_shape) == len(new_shape)
247
+ if len(new_shape) > 2:
248
+ # we only modify first two axes
249
+ assert new_shape[2:] == old_shape[2:]
250
+ # assumes first axis corresponds to output dim
251
+ if not new_shape == old_shape:
252
+ new_param = param.clone()
253
+ old_param = sd[name]
254
+ if len(new_shape) == 1:
255
+ for i in range(new_param.shape[0]):
256
+ new_param[i] = old_param[i % old_shape[0]]
257
+ elif len(new_shape) >= 2:
258
+ for i in range(new_param.shape[0]):
259
+ for j in range(new_param.shape[1]):
260
+ new_param[i, j] = old_param[i % old_shape[0], j % old_shape[1]]
261
+
262
+ n_used_old = torch.ones(old_shape[1])
263
+ for j in range(new_param.shape[1]):
264
+ n_used_old[j % old_shape[1]] += 1
265
+ n_used_new = torch.zeros(new_shape[1])
266
+ for j in range(new_param.shape[1]):
267
+ n_used_new[j] = n_used_old[j % old_shape[1]]
268
+
269
+ n_used_new = n_used_new[None, :]
270
+ while len(n_used_new.shape) < len(new_shape):
271
+ n_used_new = n_used_new.unsqueeze(-1)
272
+ new_param /= n_used_new
273
+
274
+ sd[name] = new_param
275
+
276
+ missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
277
+ sd, strict=False)
278
+ print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
279
+ if len(missing) > 0:
280
+ print(f"Missing Keys:\n {missing}")
281
+ if len(unexpected) > 0:
282
+ print(f"\nUnexpected Keys:\n {unexpected}")
283
+
284
+ def q_mean_variance(self, x_start, t):
285
+ """
286
+ Get the distribution q(x_t | x_0).
287
+ :param x_start: the [N x C x ...] tensor of noiseless inputs.
288
+ :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
289
+ :return: A tuple (mean, variance, log_variance), all of x_start's shape.
290
+ """
291
+ mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
292
+ variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
293
+ log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
294
+ return mean, variance, log_variance
295
+
296
+ def predict_start_from_noise(self, x_t, t, noise):
297
+ return (
298
+ extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
299
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
300
+ )
301
+
302
+ def predict_start_from_z_and_v(self, x_t, t, v):
303
+ # self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
304
+ # self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
305
+ return (
306
+ extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t -
307
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v
308
+ )
309
+
310
+ def predict_eps_from_z_and_v(self, x_t, t, v):
311
+ return (
312
+ extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v +
313
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t
314
+ )
315
+
316
+ def q_posterior(self, x_start, x_t, t):
317
+ posterior_mean = (
318
+ extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
319
+ extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
320
+ )
321
+ posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
322
+ posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
323
+ return posterior_mean, posterior_variance, posterior_log_variance_clipped
324
+
325
+ def p_mean_variance(self, x, t, clip_denoised: bool):
326
+ model_out = self.model(x, t)
327
+ if self.parameterization == "eps":
328
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
329
+ elif self.parameterization == "x0":
330
+ x_recon = model_out
331
+ if clip_denoised:
332
+ x_recon.clamp_(-1., 1.)
333
+
334
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
335
+ return model_mean, posterior_variance, posterior_log_variance
336
+
337
+ @torch.no_grad()
338
+ def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
339
+ b, *_, device = *x.shape, x.device
340
+ model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
341
+ noise = noise_like(x.shape, device, repeat_noise)
342
+ # no noise when t == 0
343
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
344
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
345
+
346
+ @torch.no_grad()
347
+ def p_sample_loop(self, shape, return_intermediates=False):
348
+ device = self.betas.device
349
+ b = shape[0]
350
+ img = torch.randn(shape, device=device)
351
+ intermediates = [img]
352
+ for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
353
+ img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
354
+ clip_denoised=self.clip_denoised)
355
+ if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
356
+ intermediates.append(img)
357
+ if return_intermediates:
358
+ return img, intermediates
359
+ return img
360
+
361
+ @torch.no_grad()
362
+ def sample(self, batch_size=16, return_intermediates=False):
363
+ image_size = self.image_size
364
+ channels = self.channels
365
+ return self.p_sample_loop((batch_size, channels, image_size, image_size),
366
+ return_intermediates=return_intermediates)
367
+
368
+ def q_sample(self, x_start, t, noise=None):
369
+ noise = default(noise, lambda: torch.randn_like(x_start))
370
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
371
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
372
+
373
+ def get_v(self, x, noise, t):
374
+ return (
375
+ extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise -
376
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x
377
+ )
378
+
379
+ def get_loss(self, pred, target, mean=True):
380
+ if self.loss_type == 'l1':
381
+ loss = (target - pred).abs()
382
+ if mean:
383
+ loss = loss.mean()
384
+ elif self.loss_type == 'l2':
385
+ if mean:
386
+ loss = torch.nn.functional.mse_loss(target, pred)
387
+ else:
388
+ loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
389
+ else:
390
+ raise NotImplementedError("unknown loss type '{loss_type}'")
391
+
392
+ return loss
393
+
394
+ def p_losses(self, x_start, t, noise=None):
395
+ noise = default(noise, lambda: torch.randn_like(x_start))
396
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
397
+ model_out = self.model(x_noisy, t)
398
+
399
+ loss_dict = {}
400
+ if self.parameterization == "eps":
401
+ target = noise
402
+ elif self.parameterization == "x0":
403
+ target = x_start
404
+ elif self.parameterization == "v":
405
+ target = self.get_v(x_start, noise, t)
406
+ else:
407
+ raise NotImplementedError(f"Parameterization {self.parameterization} not yet supported")
408
+
409
+ loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
410
+
411
+ log_prefix = 'train' if self.training else 'val'
412
+
413
+ loss_dict.update({f'{log_prefix}_loss_simple': loss.mean()})
414
+ loss_simple = loss.mean() * self.l_simple_weight
415
+
416
+ loss_vlb = (self.lvlb_weights[t] * loss).mean()
417
+ loss_dict.update({f'{log_prefix}_loss_vlb': loss_vlb})
418
+
419
+ loss = loss_simple + self.original_elbo_weight * loss_vlb
420
+
421
+ loss_dict.update({f'{log_prefix}_loss': loss})
422
+
423
+ return loss, loss_dict
424
+
425
+ def forward(self, x, *args, **kwargs):
426
+ # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
427
+ # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
428
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
429
+ return self.p_losses(x, t, *args, **kwargs)
430
+
431
+ def get_input(self, batch, k):
432
+ x = batch[k]
433
+ if len(x.shape) == 3:
434
+ x = x[..., None]
435
+ x = rearrange(x, 'b h w c -> b c h w')
436
+ x = x.to(memory_format=torch.contiguous_format).float()
437
+ return x
438
+
439
+ def shared_step(self, batch):
440
+ x = self.get_input(batch, self.first_stage_key)
441
+ loss, loss_dict = self(x)
442
+ return loss, loss_dict
443
+
444
+ def training_step(self, batch, batch_idx):
445
+ self.batch = batch
446
+ for k in self.ucg_training:
447
+ p = self.ucg_training[k]["p"]
448
+ val = self.ucg_training[k]["val"]
449
+ if val is None:
450
+ val = ""
451
+ for i in range(len(batch[k])):
452
+ if self.ucg_prng.choice(2, p=[1 - p, p]):
453
+ batch[k][i] = val
454
+ loss, loss_dict = self.shared_step(batch)
455
+
456
+ self.log_dict(loss_dict, prog_bar=True,
457
+ logger=True, on_step=True, on_epoch=True)
458
+
459
+ self.log("global_step", self.global_step,
460
+ prog_bar=True, logger=True, on_step=True, on_epoch=False)
461
+
462
+ if self.use_scheduler:
463
+ lr = self.optimizers().param_groups[0]['lr']
464
+ self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
465
+
466
+ return loss
467
+
468
+ @torch.no_grad()
469
+ def validation_step(self, batch, batch_idx):
470
+ _, loss_dict_no_ema = self.shared_step(batch)
471
+ with self.ema_scope():
472
+ _, loss_dict_ema = self.shared_step(batch)
473
+ loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
474
+ self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
475
+ self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
476
+
477
+ def on_train_batch_end(self, *args, **kwargs):
478
+ if self.use_ema:
479
+ self.model_ema(self.model)
480
+
481
+ def _get_rows_from_list(self, samples):
482
+ n_imgs_per_row = len(samples)
483
+ denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
484
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
485
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
486
+ return denoise_grid
487
+
488
+ @torch.no_grad()
489
+ def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
490
+ log = dict()
491
+ x = self.get_input(batch, self.first_stage_key)
492
+ N = min(x.shape[0], N)
493
+ n_row = min(x.shape[0], n_row)
494
+ x = x.to(self.device)[:N]
495
+ log["inputs"] = x
496
+
497
+ # get diffusion row
498
+ diffusion_row = list()
499
+ x_start = x[:n_row]
500
+
501
+ for t in range(self.num_timesteps):
502
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
503
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
504
+ t = t.to(self.device).long()
505
+ noise = torch.randn_like(x_start)
506
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
507
+ diffusion_row.append(x_noisy)
508
+
509
+ log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
510
+
511
+ if sample:
512
+ # get denoise row
513
+ with self.ema_scope("Plotting"):
514
+ samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
515
+
516
+ log["samples"] = samples
517
+ log["denoise_row"] = self._get_rows_from_list(denoise_row)
518
+
519
+ if return_keys:
520
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
521
+ return log
522
+ else:
523
+ return {key: log[key] for key in return_keys}
524
+ return log
525
+
526
+ def configure_optimizers(self):
527
+ lr = self.learning_rate
528
+ params = list(self.model.parameters())
529
+ if self.learn_logvar:
530
+ params = params + [self.logvar]
531
+ opt = torch.optim.AdamW(params, lr=lr)
532
+ return opt
533
+
534
+
535
+ class LatentDiffusion(DDPM):
536
+ """main class"""
537
+
538
+ def __init__(self,
539
+ first_stage_config,
540
+ cond_stage_config,
541
+ num_timesteps_cond=None,
542
+ cond_stage_key="image",
543
+ cond_stage_trainable=False,
544
+ concat_mode=True,
545
+ cond_stage_forward=None,
546
+ conditioning_key=None,
547
+ scale_factor=1.0,
548
+ scale_by_std=False,
549
+ force_null_conditioning=False,
550
+ *args, **kwargs):
551
+ self.kwargs = kwargs
552
+ self.force_null_conditioning = force_null_conditioning
553
+ self.num_timesteps_cond = default(num_timesteps_cond, 1)
554
+ self.scale_by_std = scale_by_std
555
+ self.cond_stage_trainable = cond_stage_trainable
556
+ assert self.num_timesteps_cond <= kwargs['timesteps']
557
+ if conditioning_key is None:
558
+ conditioning_key = 'concat' if concat_mode else 'crossattn'
559
+ if cond_stage_config == '__is_unconditional__' and not self.force_null_conditioning:
560
+ conditioning_key = None
561
+ ckpt_path = kwargs.pop("ckpt_path", None)
562
+ reset_ema = kwargs.pop("reset_ema", False)
563
+ reset_num_ema_updates = kwargs.pop("reset_num_ema_updates", False)
564
+ ignore_keys = kwargs.pop("ignore_keys", [])
565
+ super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
566
+ self.concat_mode = concat_mode
567
+ self.cond_stage_key = cond_stage_key
568
+ try:
569
+ self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
570
+ except:
571
+ self.num_downs = 0
572
+ if not scale_by_std:
573
+ self.scale_factor = scale_factor
574
+ else:
575
+ self.register_buffer('scale_factor', torch.tensor(scale_factor))
576
+
577
+ self.instantiate_first_stage(first_stage_config)
578
+ self.instantiate_cond_stage(cond_stage_config)
579
+ self.cond_stage_forward = cond_stage_forward
580
+ self.clip_denoised = False
581
+ self.bbox_tokenizer = None
582
+
583
+ if self.kwargs["use_imageCLIP"]:
584
+ self.proj_out = nn.Linear(1024, 768)
585
+ else:
586
+ self.proj_out = None
587
+ if self.use_pbe_weight:
588
+ print("learnable vector gene")
589
+ self.learnable_vector = nn.Parameter(torch.randn((1,1,768)), requires_grad=True)
590
+ else:
591
+ self.learnable_vector = None
592
+
593
+ if self.kwargs["use_lastzc"]: # deprecated
594
+ self.lastzc = zero_module(conv_nd(2, 4, 4, 1, 1, 0))
595
+ else:
596
+ self.lastzc = None
597
+
598
+ self.restarted_from_ckpt = False
599
+ if ckpt_path is not None:
600
+ self.init_from_ckpt(ckpt_path, ignore_keys)
601
+ self.restarted_from_ckpt = True
602
+ if reset_ema:
603
+ assert self.use_ema
604
+ print(
605
+ f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.")
606
+ self.model_ema = LitEma(self.model)
607
+ if reset_num_ema_updates:
608
+ print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ")
609
+ assert self.use_ema
610
+ self.model_ema.reset_num_updates()
611
+
612
+ def make_cond_schedule(self, ):
613
+ self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
614
+ ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
615
+ self.cond_ids[:self.num_timesteps_cond] = ids
616
+
617
+ @rank_zero_only
618
+ @torch.no_grad()
619
+ def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
620
+ # only for very first batch
621
+ if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
622
+ assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
623
+ # set rescale weight to 1./std of encodings
624
+ print("### USING STD-RESCALING ###")
625
+ x = super().get_input(batch, self.first_stage_key)
626
+ x = x.to(self.device)
627
+ encoder_posterior = self.encode_first_stage(x)
628
+ z = self.get_first_stage_encoding(encoder_posterior).detach()
629
+ del self.scale_factor
630
+ self.register_buffer('scale_factor', 1. / z.flatten().std())
631
+ print(f"setting self.scale_factor to {self.scale_factor}")
632
+ print("### USING STD-RESCALING ###")
633
+
634
+ def register_schedule(self,
635
+ given_betas=None, beta_schedule="linear", timesteps=1000,
636
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
637
+ super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
638
+
639
+ self.shorten_cond_schedule = self.num_timesteps_cond > 1
640
+ if self.shorten_cond_schedule:
641
+ self.make_cond_schedule()
642
+
643
+ def instantiate_first_stage(self, config):
644
+ model = instantiate_from_config(config)
645
+ self.first_stage_model = model.eval()
646
+ self.first_stage_model.train = disabled_train
647
+ for param in self.first_stage_model.parameters():
648
+ param.requires_grad = False
649
+
650
+ def instantiate_cond_stage(self, config):
651
+ if not self.cond_stage_trainable:
652
+ if config == "__is_first_stage__":
653
+ print("Using first stage also as cond stage.")
654
+ self.cond_stage_model = self.first_stage_model
655
+ elif config == "__is_unconditional__":
656
+ print(f"Training {self.__class__.__name__} as an unconditional model.")
657
+ self.cond_stage_model = None
658
+ else:
659
+ model = instantiate_from_config(config)
660
+ self.cond_stage_model = model
661
+ else:
662
+ assert config != '__is_first_stage__'
663
+ assert config != '__is_unconditional__'
664
+ model = instantiate_from_config(config)
665
+ self.cond_stage_model = model
666
+
667
+ def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
668
+ denoise_row = []
669
+ for zd in tqdm(samples, desc=desc):
670
+ denoise_row.append(self.decode_first_stage(zd.to(self.device),
671
+ force_not_quantize=force_no_decoder_quantization))
672
+ n_imgs_per_row = len(denoise_row)
673
+ denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W
674
+ denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
675
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
676
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
677
+ return denoise_grid
678
+
679
+ def get_first_stage_encoding(self, encoder_posterior):
680
+ if isinstance(encoder_posterior, DiagonalGaussianDistribution):
681
+ z = encoder_posterior.sample()
682
+ elif isinstance(encoder_posterior, torch.Tensor):
683
+ z = encoder_posterior
684
+ elif isinstance(encoder_posterior, AutoencoderKLOutput):
685
+ z = encoder_posterior.latent_dist.sample()
686
+ else:
687
+ raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
688
+ return self.scale_factor * z
689
+
690
+ def get_learned_conditioning(self, c):
691
+ if self.cond_stage_forward is None:
692
+ if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
693
+ c = self.cond_stage_model.encode(c)
694
+ if isinstance(c, DiagonalGaussianDistribution):
695
+ c = c.mode()
696
+ else:
697
+ c = self.cond_stage_model(c)
698
+ else:
699
+ assert hasattr(self.cond_stage_model, self.cond_stage_forward)
700
+ c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
701
+ return c
702
+
703
+ def meshgrid(self, h, w):
704
+ y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
705
+ x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
706
+
707
+ arr = torch.cat([y, x], dim=-1)
708
+ return arr
709
+
710
+ def delta_border(self, h, w):
711
+ """
712
+ :param h: height
713
+ :param w: width
714
+ :return: normalized distance to image border,
715
+ wtith min distance = 0 at border and max dist = 0.5 at image center
716
+ """
717
+ lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
718
+ arr = self.meshgrid(h, w) / lower_right_corner
719
+ dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
720
+ dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
721
+ edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
722
+ return edge_dist
723
+
724
+ def get_weighting(self, h, w, Ly, Lx, device):
725
+ weighting = self.delta_border(h, w)
726
+ weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
727
+ self.split_input_params["clip_max_weight"], )
728
+ weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
729
+
730
+ if self.split_input_params["tie_braker"]:
731
+ L_weighting = self.delta_border(Ly, Lx)
732
+ L_weighting = torch.clip(L_weighting,
733
+ self.split_input_params["clip_min_tie_weight"],
734
+ self.split_input_params["clip_max_tie_weight"])
735
+
736
+ L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
737
+ weighting = weighting * L_weighting
738
+ return weighting
739
+
740
+ def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code
741
+ """
742
+ :param x: img of size (bs, c, h, w)
743
+ :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
744
+ """
745
+ bs, nc, h, w = x.shape
746
+
747
+ # number of crops in image
748
+ Ly = (h - kernel_size[0]) // stride[0] + 1
749
+ Lx = (w - kernel_size[1]) // stride[1] + 1
750
+
751
+ if uf == 1 and df == 1:
752
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
753
+ unfold = torch.nn.Unfold(**fold_params)
754
+
755
+ fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
756
+
757
+ weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
758
+ normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap
759
+ weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
760
+
761
+ elif uf > 1 and df == 1:
762
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
763
+ unfold = torch.nn.Unfold(**fold_params)
764
+
765
+ fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
766
+ dilation=1, padding=0,
767
+ stride=(stride[0] * uf, stride[1] * uf))
768
+ fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
769
+
770
+ weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
771
+ normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap
772
+ weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
773
+
774
+ elif df > 1 and uf == 1:
775
+ fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
776
+ unfold = torch.nn.Unfold(**fold_params)
777
+
778
+ fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
779
+ dilation=1, padding=0,
780
+ stride=(stride[0] // df, stride[1] // df))
781
+ fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
782
+
783
+ weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
784
+ normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap
785
+ weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
786
+
787
+ else:
788
+ raise NotImplementedError
789
+
790
+ return fold, unfold, normalization, weighting
791
+
792
+ @torch.no_grad()
793
+ def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
794
+ cond_key=None, return_original_cond=False, bs=None, return_x=False, no_latent=False, is_controlnet=False):
795
+ x = super().get_input(batch, k)
796
+ if bs is not None:
797
+ x = x[:bs]
798
+ x = x.to(self.device)
799
+ if no_latent:
800
+ _,_,h,w = x.shape
801
+ x = resize(x, (h//8, w//8))
802
+ return [x, None]
803
+ encoder_posterior = self.encode_first_stage(x)
804
+ z = self.get_first_stage_encoding(encoder_posterior).detach()
805
+ if is_controlnet and self.lastzc is not None:
806
+ z = self.lastzc(z)
807
+
808
+ if self.model.conditioning_key is not None and not self.force_null_conditioning:
809
+ if cond_key is None:
810
+ cond_key = self.cond_stage_key
811
+ if cond_key != self.first_stage_key:
812
+ if cond_key in ['caption', 'coordinates_bbox', "txt"]:
813
+ xc = batch[cond_key]
814
+ elif cond_key in ['class_label', 'cls']:
815
+ xc = batch
816
+ else:
817
+ xc = super().get_input(batch, cond_key).to(self.device)
818
+ else:
819
+ xc = x
820
+ if not self.cond_stage_trainable or force_c_encode:
821
+ if self.kwargs["use_imageCLIP"]:
822
+ xc = resize(xc, (224,224))
823
+ xc = self.imagenet_norm((xc+1)/2)
824
+ c = xc
825
+ else:
826
+ if isinstance(xc, dict) or isinstance(xc, list):
827
+ c = self.get_learned_conditioning(xc)
828
+ else:
829
+ c = self.get_learned_conditioning(xc.to(self.device))
830
+ c = c.float()
831
+ else:
832
+ if self.kwargs["use_imageCLIP"]:
833
+ xc = resize(xc, (224,224))
834
+ xc = self.imagenet_norm((xc+1)/2)
835
+ c = xc
836
+ if bs is not None:
837
+ c = c[:bs]
838
+
839
+ if self.use_positional_encodings:
840
+ pos_x, pos_y = self.compute_latent_shifts(batch)
841
+ ckey = __conditioning_keys__[self.model.conditioning_key]
842
+ c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
843
+
844
+ else:
845
+ c = None
846
+ xc = None
847
+ if self.use_positional_encodings:
848
+ pos_x, pos_y = self.compute_latent_shifts(batch)
849
+ c = {'pos_x': pos_x, 'pos_y': pos_y}
850
+
851
+ out = [z, c]
852
+ if return_first_stage_outputs:
853
+ xrec = self.decode_first_stage(z)
854
+ out.extend([x, xrec])
855
+ if return_x:
856
+ out.extend([x])
857
+ if return_original_cond:
858
+ out.append(xc)
859
+ return out
860
+
861
+ @torch.no_grad()
862
+ def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
863
+ if predict_cids:
864
+ if z.dim() == 4:
865
+ z = torch.argmax(z.exp(), dim=1).long()
866
+ z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
867
+ z = rearrange(z, 'b h w c -> b c h w').contiguous()
868
+
869
+ z = 1. / self.scale_factor * z
870
+ output = self.first_stage_model.decode(z)
871
+ if not isinstance(output, DecoderOutput):
872
+ return output
873
+ else:
874
+ return output.sample
875
+ def decode_first_stage_train(self, z, predict_cids=False, force_not_quantize=False):
876
+ if predict_cids:
877
+ if z.dim() == 4:
878
+ z = torch.argmax(z.exp(), dim=1).long()
879
+ z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
880
+ z = rearrange(z, 'b h w c -> b c h w').contiguous()
881
+
882
+ z = 1. / self.scale_factor * z
883
+ return self.first_stage_model.decode(z)
884
+
885
+ @torch.no_grad()
886
+ def encode_first_stage(self, x):
887
+ return self.first_stage_model.encode(x)
888
+
889
+ def shared_step(self, batch, **kwargs):
890
+ x, c = self.get_input(batch, self.first_stage_key)
891
+ loss = self(x, c)
892
+ return loss
893
+
894
+ def forward(self, x, c, *args, **kwargs):
895
+ if not self.use_pbe_weight:
896
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
897
+ if self.model.conditioning_key is not None:
898
+ assert c is not None
899
+ if self.cond_stage_trainable:
900
+ c = self.get_learned_conditioning(c)
901
+ if self.shorten_cond_schedule: # TODO: drop this option
902
+ tc = self.cond_ids[t].to(self.device)
903
+ c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
904
+ return self.p_losses(x, c, t, *args, **kwargs)
905
+ # pbe negative condition
906
+ else:
907
+ t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
908
+ self.u_cond_prop=random.uniform(0, 1)
909
+ c["c_crossattn"] = [self.get_learned_conditioning(c["c_crossattn"])]
910
+ if self.u_cond_prop < self.u_cond_percent:
911
+ c["c_crossattn"] = [self.learnable_vector.repeat(x.shape[0],1,1)]
912
+ return self.p_losses(x, c, t, *args, **kwargs)
913
+
914
+
915
+ def apply_model(self, x_noisy, t, cond, return_ids=False):
916
+ if isinstance(cond, dict):
917
+ # hybrid case, cond is expected to be a dict
918
+ pass
919
+ else:
920
+ if not isinstance(cond, list):
921
+ cond = [cond]
922
+ key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
923
+ cond = {key: cond}
924
+
925
+ x_recon = self.model(x_noisy, t, **cond)
926
+
927
+ if isinstance(x_recon, tuple) and not return_ids:
928
+ return x_recon[0]
929
+ else:
930
+ return x_recon
931
+
932
+ def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
933
+ return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
934
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
935
+
936
+ def _prior_bpd(self, x_start):
937
+ """
938
+ Get the prior KL term for the variational lower-bound, measured in
939
+ bits-per-dim.
940
+ This term can't be optimized, as it only depends on the encoder.
941
+ :param x_start: the [N x C x ...] tensor of inputs.
942
+ :return: a batch of [N] KL values (in bits), one per batch element.
943
+ """
944
+ batch_size = x_start.shape[0]
945
+ t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
946
+ qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
947
+ kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
948
+ return mean_flat(kl_prior) / np.log(2.0)
949
+ def p_losses(self, x_start, cond, t, noise=None):
950
+ loss_dict = {}
951
+ noise = default(noise, lambda: torch.randn_like(x_start))
952
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
953
+ model_output, cond_output_dict = self.apply_model(x_noisy, t, cond)
954
+
955
+ prefix = 'train' if self.training else 'val'
956
+
957
+ if self.parameterization == "x0":
958
+ target = x_start
959
+ elif self.parameterization == "eps":
960
+ target = noise
961
+ elif self.parameterization == "v":
962
+ target = self.get_v(x_start, noise, t)
963
+ else:
964
+ raise NotImplementedError()
965
+ model_loss = None
966
+ if isinstance(model_output, tuple):
967
+ model_output, model_loss = model_output
968
+
969
+ if self.only_agn_simple_loss:
970
+ _, _, l_h, l_w = model_output.shape
971
+ m_agn = F.interpolate(super().get_input(self.batch, "agn_mask"), (l_h, l_w))
972
+ loss_simple = self.get_loss(model_output * (1-m_agn), target * (1-m_agn), mean=False).mean([1, 2, 3])
973
+ else:
974
+ loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
975
+ loss_dict.update({f'simple': loss_simple.mean()})
976
+
977
+ logvar_t = self.logvar[t].to(self.device)
978
+ loss = loss_simple / torch.exp(logvar_t) + logvar_t
979
+ # loss = loss_simple / torch.exp(self.logvar) + self.logvar
980
+ if self.learn_logvar:
981
+ loss_dict.update({f'gamma': loss.mean()})
982
+ loss_dict.update({'logvar': self.logvar.data.mean()})
983
+ loss = self.l_simple_weight * loss.mean()
984
+
985
+ loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
986
+ loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
987
+ if self.original_elbo_weight != 0:
988
+ loss_dict.update({f'loss_vlb': loss_vlb})
989
+ loss += (self.original_elbo_weight * loss_vlb)
990
+
991
+ if model_loss is not None:
992
+ loss += model_loss
993
+ loss_dict.update({f"model loss" : model_loss})
994
+ loss_dict.update({f'{prefix}_loss': loss})
995
+
996
+ return loss, loss_dict
997
+
998
+ def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
999
+ return_x0=False, score_corrector=None, corrector_kwargs=None):
1000
+ t_in = t
1001
+ model_out, cond_output_dict = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
1002
+ if isinstance(model_out, tuple):
1003
+ model_out, _ = model_out
1004
+
1005
+ if score_corrector is not None:
1006
+ assert self.parameterization == "eps"
1007
+ model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
1008
+
1009
+ if return_codebook_ids:
1010
+ model_out, logits = model_out
1011
+
1012
+ if self.parameterization == "eps":
1013
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
1014
+ elif self.parameterization == "x0":
1015
+ x_recon = model_out
1016
+ else:
1017
+ raise NotImplementedError()
1018
+
1019
+ if clip_denoised:
1020
+ x_recon.clamp_(-1., 1.)
1021
+ if quantize_denoised:
1022
+ x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
1023
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
1024
+ if return_codebook_ids:
1025
+ return model_mean, posterior_variance, posterior_log_variance, logits
1026
+ elif return_x0:
1027
+ return model_mean, posterior_variance, posterior_log_variance, x_recon
1028
+ else:
1029
+ return model_mean, posterior_variance, posterior_log_variance
1030
+
1031
+ @torch.no_grad()
1032
+ def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
1033
+ return_codebook_ids=False, quantize_denoised=False, return_x0=False,
1034
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
1035
+ b, *_, device = *x.shape, x.device
1036
+ outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
1037
+ return_codebook_ids=return_codebook_ids,
1038
+ quantize_denoised=quantize_denoised,
1039
+ return_x0=return_x0,
1040
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
1041
+ if return_codebook_ids:
1042
+ raise DeprecationWarning("Support dropped.")
1043
+ model_mean, _, model_log_variance, logits = outputs
1044
+ elif return_x0:
1045
+ model_mean, _, model_log_variance, x0 = outputs
1046
+ else:
1047
+ model_mean, _, model_log_variance = outputs
1048
+
1049
+ noise = noise_like(x.shape, device, repeat_noise) * temperature
1050
+ if noise_dropout > 0.:
1051
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
1052
+ # no noise when t == 0
1053
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
1054
+
1055
+ if return_codebook_ids:
1056
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
1057
+ if return_x0:
1058
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
1059
+ else:
1060
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
1061
+
1062
+ @torch.no_grad()
1063
+ def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
1064
+ img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
1065
+ score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
1066
+ log_every_t=None):
1067
+ if not log_every_t:
1068
+ log_every_t = self.log_every_t
1069
+ timesteps = self.num_timesteps
1070
+ if batch_size is not None:
1071
+ b = batch_size if batch_size is not None else shape[0]
1072
+ shape = [batch_size] + list(shape)
1073
+ else:
1074
+ b = batch_size = shape[0]
1075
+ if x_T is None:
1076
+ img = torch.randn(shape, device=self.device)
1077
+ else:
1078
+ img = x_T
1079
+ intermediates = []
1080
+ if cond is not None:
1081
+ if isinstance(cond, dict):
1082
+ cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
1083
+ list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
1084
+ else:
1085
+ cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
1086
+
1087
+ if start_T is not None:
1088
+ timesteps = min(timesteps, start_T)
1089
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
1090
+ total=timesteps) if verbose else reversed(
1091
+ range(0, timesteps))
1092
+ if type(temperature) == float:
1093
+ temperature = [temperature] * timesteps
1094
+
1095
+ for i in iterator:
1096
+ ts = torch.full((b,), i, device=self.device, dtype=torch.long)
1097
+ if self.shorten_cond_schedule:
1098
+ assert self.model.conditioning_key != 'hybrid'
1099
+ tc = self.cond_ids[ts].to(cond.device)
1100
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
1101
+
1102
+ img, x0_partial = self.p_sample(img, cond, ts,
1103
+ clip_denoised=self.clip_denoised,
1104
+ quantize_denoised=quantize_denoised, return_x0=True,
1105
+ temperature=temperature[i], noise_dropout=noise_dropout,
1106
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
1107
+ if mask is not None:
1108
+ assert x0 is not None
1109
+ img_orig = self.q_sample(x0, ts)
1110
+ img = img_orig * mask + (1. - mask) * img
1111
+
1112
+ if i % log_every_t == 0 or i == timesteps - 1:
1113
+ intermediates.append(x0_partial)
1114
+ if callback: callback(i)
1115
+ if img_callback: img_callback(img, i)
1116
+ return img, intermediates
1117
+
1118
+ @torch.no_grad()
1119
+ def p_sample_loop(self, cond, shape, return_intermediates=False,
1120
+ x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
1121
+ mask=None, x0=None, img_callback=None, start_T=None,
1122
+ log_every_t=None):
1123
+
1124
+ if not log_every_t:
1125
+ log_every_t = self.log_every_t
1126
+ device = self.betas.device
1127
+ b = shape[0]
1128
+ if x_T is None:
1129
+ img = torch.randn(shape, device=device)
1130
+ else:
1131
+ img = x_T
1132
+
1133
+ intermediates = [img]
1134
+ if timesteps is None:
1135
+ timesteps = self.num_timesteps
1136
+
1137
+ if start_T is not None:
1138
+ timesteps = min(timesteps, start_T)
1139
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
1140
+ range(0, timesteps))
1141
+
1142
+ if mask is not None:
1143
+ assert x0 is not None
1144
+ assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
1145
+
1146
+ for i in iterator:
1147
+ ts = torch.full((b,), i, device=device, dtype=torch.long)
1148
+ if self.shorten_cond_schedule:
1149
+ assert self.model.conditioning_key != 'hybrid'
1150
+ tc = self.cond_ids[ts].to(cond.device)
1151
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
1152
+
1153
+ img = self.p_sample(img, cond, ts,
1154
+ clip_denoised=self.clip_denoised,
1155
+ quantize_denoised=quantize_denoised)
1156
+ if mask is not None:
1157
+ img_orig = self.q_sample(x0, ts)
1158
+ img = img_orig * mask + (1. - mask) * img
1159
+
1160
+ if i % log_every_t == 0 or i == timesteps - 1:
1161
+ intermediates.append(img)
1162
+ if callback: callback(i)
1163
+ if img_callback: img_callback(img, i)
1164
+
1165
+ if return_intermediates:
1166
+ return img, intermediates
1167
+ return img
1168
+
1169
+ @torch.no_grad()
1170
+ def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
1171
+ verbose=True, timesteps=None, quantize_denoised=False,
1172
+ mask=None, x0=None, shape=None, **kwargs):
1173
+ if shape is None:
1174
+ shape = (batch_size, self.channels, self.image_size, self.image_size)
1175
+ if cond is not None:
1176
+ if isinstance(cond, dict):
1177
+ cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
1178
+ list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
1179
+ else:
1180
+ cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
1181
+ return self.p_sample_loop(cond,
1182
+ shape,
1183
+ return_intermediates=return_intermediates, x_T=x_T,
1184
+ verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
1185
+ mask=mask, x0=x0)
1186
+
1187
+ @torch.no_grad()
1188
+ def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs):
1189
+ if ddim:
1190
+ ddim_sampler = DDIMSampler(self)
1191
+ shape = (self.channels, self.image_size, self.image_size)
1192
+ samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size,
1193
+ shape, cond, verbose=False, **kwargs)
1194
+
1195
+ else:
1196
+ samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
1197
+ return_intermediates=True, **kwargs)
1198
+
1199
+ return samples, intermediates
1200
+
1201
+ @torch.no_grad()
1202
+ def get_unconditional_conditioning(self, batch_size, null_label=None):
1203
+ if null_label is not None:
1204
+ xc = null_label
1205
+ if isinstance(xc, ListConfig):
1206
+ xc = list(xc)
1207
+ if isinstance(xc, dict) or isinstance(xc, list):
1208
+ c = self.get_learned_conditioning(xc)
1209
+ else:
1210
+ if hasattr(xc, "to"):
1211
+ xc = xc.to(self.device)
1212
+ c = self.get_learned_conditioning(xc)
1213
+ else:
1214
+ if self.cond_stage_key in ["class_label", "cls"]:
1215
+ xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device)
1216
+ return self.get_learned_conditioning(xc)
1217
+ else:
1218
+ raise NotImplementedError("todo")
1219
+ if isinstance(c, list): # in case the encoder gives us a list
1220
+ for i in range(len(c)):
1221
+ c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device)
1222
+ else:
1223
+ c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device)
1224
+ return c
1225
+
1226
+ @torch.no_grad()
1227
+ def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=50, ddim_eta=0., return_keys=None,
1228
+ quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
1229
+ plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
1230
+ use_ema_scope=True,
1231
+ **kwargs):
1232
+ ema_scope = self.ema_scope if use_ema_scope else nullcontext
1233
+ use_ddim = ddim_steps is not None
1234
+
1235
+ log = dict()
1236
+ z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
1237
+ return_first_stage_outputs=True,
1238
+ force_c_encode=True,
1239
+ return_original_cond=True,
1240
+ bs=N)
1241
+ N = min(x.shape[0], N)
1242
+ n_row = min(x.shape[0], n_row)
1243
+ log["inputs"] = x
1244
+ log["reconstruction"] = xrec
1245
+ if self.model.conditioning_key is not None:
1246
+ if hasattr(self.cond_stage_model, "decode"):
1247
+ xc = self.cond_stage_model.decode(c)
1248
+ log["conditioning"] = xc
1249
+ elif self.cond_stage_key in ["caption", "txt"]:
1250
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
1251
+ log["conditioning"] = xc
1252
+ elif self.cond_stage_key in ['class_label', "cls"]:
1253
+ try:
1254
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
1255
+ log['conditioning'] = xc
1256
+ except KeyError:
1257
+ # probably no "human_label" in batch
1258
+ pass
1259
+ elif isimage(xc):
1260
+ log["conditioning"] = xc
1261
+ if ismap(xc):
1262
+ log["original_conditioning"] = self.to_rgb(xc)
1263
+
1264
+ if plot_diffusion_rows:
1265
+ # get diffusion row
1266
+ diffusion_row = list()
1267
+ z_start = z[:n_row]
1268
+ for t in range(self.num_timesteps):
1269
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
1270
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
1271
+ t = t.to(self.device).long()
1272
+ noise = torch.randn_like(z_start)
1273
+ z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
1274
+ diffusion_row.append(self.decode_first_stage(z_noisy))
1275
+
1276
+ diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
1277
+ diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
1278
+ diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
1279
+ diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
1280
+ log["diffusion_row"] = diffusion_grid
1281
+
1282
+ if sample:
1283
+ # get denoise row
1284
+ with ema_scope("Sampling"):
1285
+ samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
1286
+ ddim_steps=ddim_steps, eta=ddim_eta)
1287
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
1288
+ x_samples = self.decode_first_stage(samples)
1289
+ log["samples"] = x_samples
1290
+ if plot_denoise_rows:
1291
+ denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
1292
+ log["denoise_row"] = denoise_grid
1293
+
1294
+ if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
1295
+ self.first_stage_model, IdentityFirstStage):
1296
+ # also display when quantizing x0 while sampling
1297
+ with ema_scope("Plotting Quantized Denoised"):
1298
+ samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
1299
+ ddim_steps=ddim_steps, eta=ddim_eta,
1300
+ quantize_denoised=True)
1301
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
1302
+ # quantize_denoised=True)
1303
+ x_samples = self.decode_first_stage(samples.to(self.device))
1304
+ log["samples_x0_quantized"] = x_samples
1305
+
1306
+ if unconditional_guidance_scale > 1.0:
1307
+ uc = self.get_unconditional_conditioning(N, unconditional_guidance_label)
1308
+ if self.model.conditioning_key == "crossattn-adm":
1309
+ uc = {"c_crossattn": [uc], "c_adm": c["c_adm"]}
1310
+ with ema_scope("Sampling with classifier-free guidance"):
1311
+ samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
1312
+ ddim_steps=ddim_steps, eta=ddim_eta,
1313
+ unconditional_guidance_scale=unconditional_guidance_scale,
1314
+ unconditional_conditioning=uc,
1315
+ )
1316
+ x_samples_cfg = self.decode_first_stage(samples_cfg)
1317
+ log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
1318
+
1319
+ if inpaint:
1320
+ # make a simple center square
1321
+ b, h, w = z.shape[0], z.shape[2], z.shape[3]
1322
+ mask = torch.ones(N, h, w).to(self.device)
1323
+ # zeros will be filled in
1324
+ mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
1325
+ mask = mask[:, None, ...]
1326
+ with ema_scope("Plotting Inpaint"):
1327
+ samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta,
1328
+ ddim_steps=ddim_steps, x0=z[:N], mask=mask)
1329
+ x_samples = self.decode_first_stage(samples.to(self.device))
1330
+ log["samples_inpainting"] = x_samples
1331
+ log["mask"] = mask
1332
+
1333
+ # outpaint
1334
+ mask = 1. - mask
1335
+ with ema_scope("Plotting Outpaint"):
1336
+ samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta,
1337
+ ddim_steps=ddim_steps, x0=z[:N], mask=mask)
1338
+ x_samples = self.decode_first_stage(samples.to(self.device))
1339
+ log["samples_outpainting"] = x_samples
1340
+
1341
+ if plot_progressive_rows:
1342
+ with ema_scope("Plotting Progressives"):
1343
+ img, progressives = self.progressive_denoising(c,
1344
+ shape=(self.channels, self.image_size, self.image_size),
1345
+ batch_size=N)
1346
+ prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
1347
+ log["progressive_row"] = prog_row
1348
+
1349
+ if return_keys:
1350
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
1351
+ return log
1352
+ else:
1353
+ return {key: log[key] for key in return_keys}
1354
+ return log
1355
+
1356
+ def configure_optimizers(self):
1357
+ lr = self.learning_rate
1358
+ params = list(self.model.parameters())
1359
+ if self.cond_stage_trainable:
1360
+ print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
1361
+ params = params + list(self.cond_stage_model.parameters())
1362
+ if self.learn_logvar:
1363
+ print('Diffusion model optimizing logvar')
1364
+ params.append(self.logvar)
1365
+ opt = torch.optim.AdamW(params, lr=lr)
1366
+ if self.use_scheduler:
1367
+ assert 'target' in self.scheduler_config
1368
+ scheduler = instantiate_from_config(self.scheduler_config)
1369
+
1370
+ print("Setting up LambdaLR scheduler...")
1371
+ scheduler = [
1372
+ {
1373
+ 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
1374
+ 'interval': 'step',
1375
+ 'frequency': 1
1376
+ }]
1377
+ return [opt], scheduler
1378
+ return opt
1379
+
1380
+ @torch.no_grad()
1381
+ def to_rgb(self, x):
1382
+ x = x.float()
1383
+ if not hasattr(self, "colorize"):
1384
+ self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
1385
+ x = nn.functional.conv2d(x, weight=self.colorize)
1386
+ x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
1387
+ return x
1388
+
1389
+
1390
+ class DiffusionWrapper(pl.LightningModule):
1391
+ def __init__(self, diff_model_config, conditioning_key):
1392
+ super().__init__()
1393
+ self.sequential_cross_attn = diff_model_config.pop("sequential_crossattn", False)
1394
+ self.diffusion_model = instantiate_from_config(diff_model_config)
1395
+ self.conditioning_key = conditioning_key
1396
+ assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm', 'crossattn-adm']
1397
+
1398
+ def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None):
1399
+ if self.conditioning_key is None:
1400
+ out = self.diffusion_model(x, t)
1401
+ elif self.conditioning_key == 'concat':
1402
+ xc = torch.cat([x] + c_concat, dim=1)
1403
+ out = self.diffusion_model(xc, t)
1404
+ elif self.conditioning_key == 'crossattn':
1405
+ if not self.sequential_cross_attn:
1406
+ cc = torch.cat(c_crossattn, 1)
1407
+ else:
1408
+ cc = c_crossattn
1409
+ out = self.diffusion_model(x, t, context=cc)
1410
+ elif self.conditioning_key == 'hybrid':
1411
+ xc = torch.cat([x] + c_concat, dim=1)
1412
+ cc = torch.cat(c_crossattn, 1)
1413
+ out = self.diffusion_model(xc, t, context=cc)
1414
+ elif self.conditioning_key == 'hybrid-adm':
1415
+ assert c_adm is not None
1416
+ xc = torch.cat([x] + c_concat, dim=1)
1417
+ cc = torch.cat(c_crossattn, 1)
1418
+ out = self.diffusion_model(xc, t, context=cc, y=c_adm)
1419
+ elif self.conditioning_key == 'crossattn-adm':
1420
+ assert c_adm is not None
1421
+ cc = torch.cat(c_crossattn, 1)
1422
+ out = self.diffusion_model(x, t, context=cc, y=c_adm)
1423
+ elif self.conditioning_key == 'adm':
1424
+ cc = c_crossattn[0]
1425
+ out = self.diffusion_model(x, t, y=cc)
1426
+ else:
1427
+ raise NotImplementedError()
1428
+
1429
+ return out
1430
+
1431
+
1432
+ class LatentUpscaleDiffusion(LatentDiffusion):
1433
+ def __init__(self, *args, low_scale_config, low_scale_key="LR", noise_level_key=None, **kwargs):
1434
+ super().__init__(*args, **kwargs)
1435
+ # assumes that neither the cond_stage nor the low_scale_model contain trainable params
1436
+ assert not self.cond_stage_trainable
1437
+ self.instantiate_low_stage(low_scale_config)
1438
+ self.low_scale_key = low_scale_key
1439
+ self.noise_level_key = noise_level_key
1440
+
1441
+ def instantiate_low_stage(self, config):
1442
+ model = instantiate_from_config(config)
1443
+ self.low_scale_model = model.eval()
1444
+ self.low_scale_model.train = disabled_train
1445
+ for param in self.low_scale_model.parameters():
1446
+ param.requires_grad = False
1447
+
1448
+ @torch.no_grad()
1449
+ def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False):
1450
+ if not log_mode:
1451
+ z, c = super().get_input(batch, k, force_c_encode=True, bs=bs)
1452
+ else:
1453
+ z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
1454
+ force_c_encode=True, return_original_cond=True, bs=bs)
1455
+ x_low = batch[self.low_scale_key][:bs]
1456
+ x_low = rearrange(x_low, 'b h w c -> b c h w')
1457
+ x_low = x_low.to(memory_format=torch.contiguous_format).float()
1458
+ zx, noise_level = self.low_scale_model(x_low)
1459
+ if self.noise_level_key is not None:
1460
+ # get noise level from batch instead, e.g. when extracting a custom noise level for bsr
1461
+ raise NotImplementedError('TODO')
1462
+
1463
+ all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level}
1464
+ if log_mode:
1465
+ # TODO: maybe disable if too expensive
1466
+ x_low_rec = self.low_scale_model.decode(zx)
1467
+ return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level
1468
+ return z, all_conds
1469
+
1470
+ @torch.no_grad()
1471
+ def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
1472
+ plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True,
1473
+ unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True,
1474
+ **kwargs):
1475
+ ema_scope = self.ema_scope if use_ema_scope else nullcontext
1476
+ use_ddim = ddim_steps is not None
1477
+
1478
+ log = dict()
1479
+ z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N,
1480
+ log_mode=True)
1481
+ N = min(x.shape[0], N)
1482
+ n_row = min(x.shape[0], n_row)
1483
+ log["inputs"] = x
1484
+ log["reconstruction"] = xrec
1485
+ log["x_lr"] = x_low
1486
+ log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec
1487
+ if self.model.conditioning_key is not None:
1488
+ if hasattr(self.cond_stage_model, "decode"):
1489
+ xc = self.cond_stage_model.decode(c)
1490
+ log["conditioning"] = xc
1491
+ elif self.cond_stage_key in ["caption", "txt"]:
1492
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
1493
+ log["conditioning"] = xc
1494
+ elif self.cond_stage_key in ['class_label', 'cls']:
1495
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
1496
+ log['conditioning'] = xc
1497
+ elif isimage(xc):
1498
+ log["conditioning"] = xc
1499
+ if ismap(xc):
1500
+ log["original_conditioning"] = self.to_rgb(xc)
1501
+
1502
+ if plot_diffusion_rows:
1503
+ # get diffusion row
1504
+ diffusion_row = list()
1505
+ z_start = z[:n_row]
1506
+ for t in range(self.num_timesteps):
1507
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
1508
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
1509
+ t = t.to(self.device).long()
1510
+ noise = torch.randn_like(z_start)
1511
+ z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
1512
+ diffusion_row.append(self.decode_first_stage(z_noisy))
1513
+
1514
+ diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
1515
+ diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
1516
+ diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
1517
+ diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
1518
+ log["diffusion_row"] = diffusion_grid
1519
+
1520
+ if sample:
1521
+ # get denoise row
1522
+ with ema_scope("Sampling"):
1523
+ samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
1524
+ ddim_steps=ddim_steps, eta=ddim_eta)
1525
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
1526
+ x_samples = self.decode_first_stage(samples)
1527
+ log["samples"] = x_samples
1528
+ if plot_denoise_rows:
1529
+ denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
1530
+ log["denoise_row"] = denoise_grid
1531
+
1532
+ if unconditional_guidance_scale > 1.0:
1533
+ uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label)
1534
+ # TODO explore better "unconditional" choices for the other keys
1535
+ # maybe guide away from empty text label and highest noise level and maximally degraded zx?
1536
+ uc = dict()
1537
+ for k in c:
1538
+ if k == "c_crossattn":
1539
+ assert isinstance(c[k], list) and len(c[k]) == 1
1540
+ uc[k] = [uc_tmp]
1541
+ elif k == "c_adm": # todo: only run with text-based guidance?
1542
+ assert isinstance(c[k], torch.Tensor)
1543
+ #uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level
1544
+ uc[k] = c[k]
1545
+ elif isinstance(c[k], list):
1546
+ uc[k] = [c[k][i] for i in range(len(c[k]))]
1547
+ else:
1548
+ uc[k] = c[k]
1549
+
1550
+ with ema_scope("Sampling with classifier-free guidance"):
1551
+ samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
1552
+ ddim_steps=ddim_steps, eta=ddim_eta,
1553
+ unconditional_guidance_scale=unconditional_guidance_scale,
1554
+ unconditional_conditioning=uc,
1555
+ )
1556
+ x_samples_cfg = self.decode_first_stage(samples_cfg)
1557
+ log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
1558
+
1559
+ if plot_progressive_rows:
1560
+ with ema_scope("Plotting Progressives"):
1561
+ img, progressives = self.progressive_denoising(c,
1562
+ shape=(self.channels, self.image_size, self.image_size),
1563
+ batch_size=N)
1564
+ prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
1565
+ log["progressive_row"] = prog_row
1566
+
1567
+ return log
1568
+
1569
+
1570
+ class LatentFinetuneDiffusion(LatentDiffusion):
1571
+ """
1572
+ Basis for different finetunas, such as inpainting or depth2image
1573
+ To disable finetuning mode, set finetune_keys to None
1574
+ """
1575
+
1576
+ def __init__(self,
1577
+ concat_keys: tuple,
1578
+ finetune_keys=("model.diffusion_model.input_blocks.0.0.weight",
1579
+ "model_ema.diffusion_modelinput_blocks00weight"
1580
+ ),
1581
+ keep_finetune_dims=4,
1582
+ # if model was trained without concat mode before and we would like to keep these channels
1583
+ c_concat_log_start=None, # to log reconstruction of c_concat codes
1584
+ c_concat_log_end=None,
1585
+ *args, **kwargs
1586
+ ):
1587
+ ckpt_path = kwargs.pop("ckpt_path", None)
1588
+ ignore_keys = kwargs.pop("ignore_keys", list())
1589
+ super().__init__(*args, **kwargs)
1590
+ self.finetune_keys = finetune_keys
1591
+ self.concat_keys = concat_keys
1592
+ self.keep_dims = keep_finetune_dims
1593
+ self.c_concat_log_start = c_concat_log_start
1594
+ self.c_concat_log_end = c_concat_log_end
1595
+ if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint'
1596
+ if exists(ckpt_path):
1597
+ self.init_from_ckpt(ckpt_path, ignore_keys)
1598
+
1599
+ def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
1600
+ sd = torch.load(path, map_location="cpu")
1601
+ if "state_dict" in list(sd.keys()):
1602
+ sd = sd["state_dict"]
1603
+ keys = list(sd.keys())
1604
+ for k in keys:
1605
+ for ik in ignore_keys:
1606
+ if k.startswith(ik):
1607
+ print("Deleting key {} from state_dict.".format(k))
1608
+ del sd[k]
1609
+
1610
+ # make it explicit, finetune by including extra input channels
1611
+ if exists(self.finetune_keys) and k in self.finetune_keys:
1612
+ new_entry = None
1613
+ for name, param in self.named_parameters():
1614
+ if name in self.finetune_keys:
1615
+ print(
1616
+ f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only")
1617
+ new_entry = torch.zeros_like(param) # zero init
1618
+ assert exists(new_entry), 'did not find matching parameter to modify'
1619
+ new_entry[:, :self.keep_dims, ...] = sd[k]
1620
+ sd[k] = new_entry
1621
+
1622
+ missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
1623
+ sd, strict=False)
1624
+ print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
1625
+ if len(missing) > 0:
1626
+ print(f"Missing Keys: {missing}")
1627
+ if len(unexpected) > 0:
1628
+ print(f"Unexpected Keys: {unexpected}")
1629
+
1630
+ @torch.no_grad()
1631
+ def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
1632
+ quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
1633
+ plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
1634
+ use_ema_scope=True,
1635
+ **kwargs):
1636
+ ema_scope = self.ema_scope if use_ema_scope else nullcontext
1637
+ use_ddim = ddim_steps is not None
1638
+
1639
+ log = dict()
1640
+ z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True)
1641
+ c_cat, c = c["c_concat"][0], c["c_crossattn"][0]
1642
+ N = min(x.shape[0], N)
1643
+ n_row = min(x.shape[0], n_row)
1644
+ log["inputs"] = x
1645
+ log["reconstruction"] = xrec
1646
+ if self.model.conditioning_key is not None:
1647
+ if hasattr(self.cond_stage_model, "decode"):
1648
+ xc = self.cond_stage_model.decode(c)
1649
+ log["conditioning"] = xc
1650
+ elif self.cond_stage_key in ["caption", "txt"]:
1651
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
1652
+ log["conditioning"] = xc
1653
+ elif self.cond_stage_key in ['class_label', 'cls']:
1654
+ xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
1655
+ log['conditioning'] = xc
1656
+ elif isimage(xc):
1657
+ log["conditioning"] = xc
1658
+ if ismap(xc):
1659
+ log["original_conditioning"] = self.to_rgb(xc)
1660
+
1661
+ if not (self.c_concat_log_start is None and self.c_concat_log_end is None):
1662
+ log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end])
1663
+
1664
+ if plot_diffusion_rows:
1665
+ # get diffusion row
1666
+ diffusion_row = list()
1667
+ z_start = z[:n_row]
1668
+ for t in range(self.num_timesteps):
1669
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
1670
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
1671
+ t = t.to(self.device).long()
1672
+ noise = torch.randn_like(z_start)
1673
+ z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
1674
+ diffusion_row.append(self.decode_first_stage(z_noisy))
1675
+
1676
+ diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
1677
+ diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
1678
+ diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
1679
+ diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
1680
+ log["diffusion_row"] = diffusion_grid
1681
+
1682
+ if sample:
1683
+ # get denoise row
1684
+ with ema_scope("Sampling"):
1685
+ samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
1686
+ batch_size=N, ddim=use_ddim,
1687
+ ddim_steps=ddim_steps, eta=ddim_eta)
1688
+ # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
1689
+ x_samples = self.decode_first_stage(samples)
1690
+ log["samples"] = x_samples
1691
+ if plot_denoise_rows:
1692
+ denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
1693
+ log["denoise_row"] = denoise_grid
1694
+
1695
+ if unconditional_guidance_scale > 1.0:
1696
+ uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label)
1697
+ uc_cat = c_cat
1698
+ uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]}
1699
+ with ema_scope("Sampling with classifier-free guidance"):
1700
+ samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
1701
+ batch_size=N, ddim=use_ddim,
1702
+ ddim_steps=ddim_steps, eta=ddim_eta,
1703
+ unconditional_guidance_scale=unconditional_guidance_scale,
1704
+ unconditional_conditioning=uc_full,
1705
+ )
1706
+ x_samples_cfg = self.decode_first_stage(samples_cfg)
1707
+ log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
1708
+
1709
+ return log
1710
+
1711
+
1712
+ class LatentInpaintDiffusion(LatentFinetuneDiffusion):
1713
+ """
1714
+ can either run as pure inpainting model (only concat mode) or with mixed conditionings,
1715
+ e.g. mask as concat and text via cross-attn.
1716
+ To disable finetuning mode, set finetune_keys to None
1717
+ """
1718
+
1719
+ def __init__(self,
1720
+ concat_keys=("mask", "masked_image"),
1721
+ masked_image_key="masked_image",
1722
+ *args, **kwargs
1723
+ ):
1724
+ super().__init__(concat_keys, *args, **kwargs)
1725
+ self.masked_image_key = masked_image_key
1726
+ assert self.masked_image_key in concat_keys
1727
+
1728
+ @torch.no_grad()
1729
+ def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
1730
+ # note: restricted to non-trainable encoders currently
1731
+ assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting'
1732
+ z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
1733
+ force_c_encode=True, return_original_cond=True, bs=bs)
1734
+
1735
+ assert exists(self.concat_keys)
1736
+ c_cat = list()
1737
+ for ck in self.concat_keys:
1738
+ cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
1739
+ if bs is not None:
1740
+ cc = cc[:bs]
1741
+ cc = cc.to(self.device)
1742
+ bchw = z.shape
1743
+ if ck != self.masked_image_key:
1744
+ cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
1745
+ else:
1746
+ cc = self.get_first_stage_encoding(self.encode_first_stage(cc))
1747
+ c_cat.append(cc)
1748
+ c_cat = torch.cat(c_cat, dim=1)
1749
+ all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
1750
+ if return_first_stage_outputs:
1751
+ return z, all_conds, x, xrec, xc
1752
+ return z, all_conds
1753
+
1754
+ @torch.no_grad()
1755
+ def log_images(self, *args, **kwargs):
1756
+ log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs)
1757
+ log["masked_image"] = rearrange(args[0]["masked_image"],
1758
+ 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
1759
+ return log
1760
+
1761
+
1762
+ class LatentDepth2ImageDiffusion(LatentFinetuneDiffusion):
1763
+ """
1764
+ condition on monocular depth estimation
1765
+ """
1766
+
1767
+ def __init__(self, depth_stage_config, concat_keys=("midas_in",), *args, **kwargs):
1768
+ super().__init__(concat_keys=concat_keys, *args, **kwargs)
1769
+ self.depth_model = instantiate_from_config(depth_stage_config)
1770
+ self.depth_stage_key = concat_keys[0]
1771
+
1772
+ @torch.no_grad()
1773
+ def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
1774
+ # note: restricted to non-trainable encoders currently
1775
+ assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for depth2img'
1776
+ z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
1777
+ force_c_encode=True, return_original_cond=True, bs=bs)
1778
+
1779
+ assert exists(self.concat_keys)
1780
+ assert len(self.concat_keys) == 1
1781
+ c_cat = list()
1782
+ for ck in self.concat_keys:
1783
+ cc = batch[ck]
1784
+ if bs is not None:
1785
+ cc = cc[:bs]
1786
+ cc = cc.to(self.device)
1787
+ cc = self.depth_model(cc)
1788
+ cc = torch.nn.functional.interpolate(
1789
+ cc,
1790
+ size=z.shape[2:],
1791
+ mode="bicubic",
1792
+ align_corners=False,
1793
+ )
1794
+
1795
+ depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3],
1796
+ keepdim=True)
1797
+ cc = 2. * (cc - depth_min) / (depth_max - depth_min + 0.001) - 1.
1798
+ c_cat.append(cc)
1799
+ c_cat = torch.cat(c_cat, dim=1)
1800
+ all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
1801
+ if return_first_stage_outputs:
1802
+ return z, all_conds, x, xrec, xc
1803
+ return z, all_conds
1804
+
1805
+ @torch.no_grad()
1806
+ def log_images(self, *args, **kwargs):
1807
+ log = super().log_images(*args, **kwargs)
1808
+ depth = self.depth_model(args[0][self.depth_stage_key])
1809
+ depth_min, depth_max = torch.amin(depth, dim=[1, 2, 3], keepdim=True), \
1810
+ torch.amax(depth, dim=[1, 2, 3], keepdim=True)
1811
+ log["depth"] = 2. * (depth - depth_min) / (depth_max - depth_min) - 1.
1812
+ return log
1813
+
1814
+
1815
+ class LatentUpscaleFinetuneDiffusion(LatentFinetuneDiffusion):
1816
+ """
1817
+ condition on low-res image (and optionally on some spatial noise augmentation)
1818
+ """
1819
+ def __init__(self, concat_keys=("lr",), reshuffle_patch_size=None,
1820
+ low_scale_config=None, low_scale_key=None, *args, **kwargs):
1821
+ super().__init__(concat_keys=concat_keys, *args, **kwargs)
1822
+ self.reshuffle_patch_size = reshuffle_patch_size
1823
+ self.low_scale_model = None
1824
+ if low_scale_config is not None:
1825
+ print("Initializing a low-scale model")
1826
+ assert exists(low_scale_key)
1827
+ self.instantiate_low_stage(low_scale_config)
1828
+ self.low_scale_key = low_scale_key
1829
+
1830
+ def instantiate_low_stage(self, config):
1831
+ model = instantiate_from_config(config)
1832
+ self.low_scale_model = model.eval()
1833
+ self.low_scale_model.train = disabled_train
1834
+ for param in self.low_scale_model.parameters():
1835
+ param.requires_grad = False
1836
+
1837
+ @torch.no_grad()
1838
+ def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
1839
+ # note: restricted to non-trainable encoders currently
1840
+ assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for upscaling-ft'
1841
+ z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
1842
+ force_c_encode=True, return_original_cond=True, bs=bs)
1843
+
1844
+ assert exists(self.concat_keys)
1845
+ assert len(self.concat_keys) == 1
1846
+ # optionally make spatial noise_level here
1847
+ c_cat = list()
1848
+ noise_level = None
1849
+ for ck in self.concat_keys:
1850
+ cc = batch[ck]
1851
+ cc = rearrange(cc, 'b h w c -> b c h w')
1852
+ if exists(self.reshuffle_patch_size):
1853
+ assert isinstance(self.reshuffle_patch_size, int)
1854
+ cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w',
1855
+ p1=self.reshuffle_patch_size, p2=self.reshuffle_patch_size)
1856
+ if bs is not None:
1857
+ cc = cc[:bs]
1858
+ cc = cc.to(self.device)
1859
+ if exists(self.low_scale_model) and ck == self.low_scale_key:
1860
+ cc, noise_level = self.low_scale_model(cc)
1861
+ c_cat.append(cc)
1862
+ c_cat = torch.cat(c_cat, dim=1)
1863
+ if exists(noise_level):
1864
+ all_conds = {"c_concat": [c_cat], "c_crossattn": [c], "c_adm": noise_level}
1865
+ else:
1866
+ all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
1867
+ if return_first_stage_outputs:
1868
+ return z, all_conds, x, xrec, xc
1869
+ return z, all_conds
1870
+
1871
+ @torch.no_grad()
1872
+ def log_images(self, *args, **kwargs):
1873
+ log = super().log_images(*args, **kwargs)
1874
+ log["lr"] = rearrange(args[0]["lr"], 'b h w c -> b c h w')
1875
+ return log
ldm/models/diffusion/dpm_solver/__init__.py ADDED
@@ -0,0 +1 @@
 
 
1
+ from .sampler import DPMSolverSampler
ldm/models/diffusion/dpm_solver/dpm_solver.py ADDED
@@ -0,0 +1,1154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn.functional as F
3
+ import math
4
+ from tqdm import tqdm
5
+
6
+
7
+ class NoiseScheduleVP:
8
+ def __init__(
9
+ self,
10
+ schedule='discrete',
11
+ betas=None,
12
+ alphas_cumprod=None,
13
+ continuous_beta_0=0.1,
14
+ continuous_beta_1=20.,
15
+ ):
16
+ """Create a wrapper class for the forward SDE (VP type).
17
+ ***
18
+ Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
19
+ We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
20
+ ***
21
+ The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
22
+ We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
23
+ Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
24
+ log_alpha_t = self.marginal_log_mean_coeff(t)
25
+ sigma_t = self.marginal_std(t)
26
+ lambda_t = self.marginal_lambda(t)
27
+ Moreover, as lambda(t) is an invertible function, we also support its inverse function:
28
+ t = self.inverse_lambda(lambda_t)
29
+ ===============================================================
30
+ We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
31
+ 1. For discrete-time DPMs:
32
+ For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
33
+ t_i = (i + 1) / N
34
+ e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
35
+ We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
36
+ Args:
37
+ betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
38
+ alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
39
+ Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
40
+ **Important**: Please pay special attention for the args for `alphas_cumprod`:
41
+ The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
42
+ q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
43
+ Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
44
+ alpha_{t_n} = \sqrt{\hat{alpha_n}},
45
+ and
46
+ log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
47
+ 2. For continuous-time DPMs:
48
+ We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
49
+ schedule are the default settings in DDPM and improved-DDPM:
50
+ Args:
51
+ beta_min: A `float` number. The smallest beta for the linear schedule.
52
+ beta_max: A `float` number. The largest beta for the linear schedule.
53
+ cosine_s: A `float` number. The hyperparameter in the cosine schedule.
54
+ cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
55
+ T: A `float` number. The ending time of the forward process.
56
+ ===============================================================
57
+ Args:
58
+ schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
59
+ 'linear' or 'cosine' for continuous-time DPMs.
60
+ Returns:
61
+ A wrapper object of the forward SDE (VP type).
62
+
63
+ ===============================================================
64
+ Example:
65
+ # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
66
+ >>> ns = NoiseScheduleVP('discrete', betas=betas)
67
+ # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
68
+ >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
69
+ # For continuous-time DPMs (VPSDE), linear schedule:
70
+ >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
71
+ """
72
+
73
+ if schedule not in ['discrete', 'linear', 'cosine']:
74
+ raise ValueError(
75
+ "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
76
+ schedule))
77
+
78
+ self.schedule = schedule
79
+ if schedule == 'discrete':
80
+ if betas is not None:
81
+ log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
82
+ else:
83
+ assert alphas_cumprod is not None
84
+ log_alphas = 0.5 * torch.log(alphas_cumprod)
85
+ self.total_N = len(log_alphas)
86
+ self.T = 1.
87
+ self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1))
88
+ self.log_alpha_array = log_alphas.reshape((1, -1,))
89
+ else:
90
+ self.total_N = 1000
91
+ self.beta_0 = continuous_beta_0
92
+ self.beta_1 = continuous_beta_1
93
+ self.cosine_s = 0.008
94
+ self.cosine_beta_max = 999.
95
+ self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * (
96
+ 1. + self.cosine_s) / math.pi - self.cosine_s
97
+ self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.))
98
+ self.schedule = schedule
99
+ if schedule == 'cosine':
100
+ # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
101
+ # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
102
+ self.T = 0.9946
103
+ else:
104
+ self.T = 1.
105
+
106
+ def marginal_log_mean_coeff(self, t):
107
+ """
108
+ Compute log(alpha_t) of a given continuous-time label t in [0, T].
109
+ """
110
+ if self.schedule == 'discrete':
111
+ return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device),
112
+ self.log_alpha_array.to(t.device)).reshape((-1))
113
+ elif self.schedule == 'linear':
114
+ return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
115
+ elif self.schedule == 'cosine':
116
+ log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.))
117
+ log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
118
+ return log_alpha_t
119
+
120
+ def marginal_alpha(self, t):
121
+ """
122
+ Compute alpha_t of a given continuous-time label t in [0, T].
123
+ """
124
+ return torch.exp(self.marginal_log_mean_coeff(t))
125
+
126
+ def marginal_std(self, t):
127
+ """
128
+ Compute sigma_t of a given continuous-time label t in [0, T].
129
+ """
130
+ return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
131
+
132
+ def marginal_lambda(self, t):
133
+ """
134
+ Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
135
+ """
136
+ log_mean_coeff = self.marginal_log_mean_coeff(t)
137
+ log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
138
+ return log_mean_coeff - log_std
139
+
140
+ def inverse_lambda(self, lamb):
141
+ """
142
+ Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
143
+ """
144
+ if self.schedule == 'linear':
145
+ tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
146
+ Delta = self.beta_0 ** 2 + tmp
147
+ return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
148
+ elif self.schedule == 'discrete':
149
+ log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb)
150
+ t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]),
151
+ torch.flip(self.t_array.to(lamb.device), [1]))
152
+ return t.reshape((-1,))
153
+ else:
154
+ log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
155
+ t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * (
156
+ 1. + self.cosine_s) / math.pi - self.cosine_s
157
+ t = t_fn(log_alpha)
158
+ return t
159
+
160
+
161
+ def model_wrapper(
162
+ model,
163
+ noise_schedule,
164
+ model_type="noise",
165
+ model_kwargs={},
166
+ guidance_type="uncond",
167
+ condition=None,
168
+ unconditional_condition=None,
169
+ guidance_scale=1.,
170
+ classifier_fn=None,
171
+ classifier_kwargs={},
172
+ ):
173
+ """Create a wrapper function for the noise prediction model.
174
+ DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
175
+ firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
176
+ We support four types of the diffusion model by setting `model_type`:
177
+ 1. "noise": noise prediction model. (Trained by predicting noise).
178
+ 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
179
+ 3. "v": velocity prediction model. (Trained by predicting the velocity).
180
+ The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
181
+ [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
182
+ arXiv preprint arXiv:2202.00512 (2022).
183
+ [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
184
+ arXiv preprint arXiv:2210.02303 (2022).
185
+
186
+ 4. "score": marginal score function. (Trained by denoising score matching).
187
+ Note that the score function and the noise prediction model follows a simple relationship:
188
+ ```
189
+ noise(x_t, t) = -sigma_t * score(x_t, t)
190
+ ```
191
+ We support three types of guided sampling by DPMs by setting `guidance_type`:
192
+ 1. "uncond": unconditional sampling by DPMs.
193
+ The input `model` has the following format:
194
+ ``
195
+ model(x, t_input, **model_kwargs) -> noise | x_start | v | score
196
+ ``
197
+ 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
198
+ The input `model` has the following format:
199
+ ``
200
+ model(x, t_input, **model_kwargs) -> noise | x_start | v | score
201
+ ``
202
+ The input `classifier_fn` has the following format:
203
+ ``
204
+ classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
205
+ ``
206
+ [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
207
+ in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
208
+ 3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
209
+ The input `model` has the following format:
210
+ ``
211
+ model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
212
+ ``
213
+ And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
214
+ [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
215
+ arXiv preprint arXiv:2207.12598 (2022).
216
+
217
+ The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
218
+ or continuous-time labels (i.e. epsilon to T).
219
+ We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
220
+ ``
221
+ def model_fn(x, t_continuous) -> noise:
222
+ t_input = get_model_input_time(t_continuous)
223
+ return noise_pred(model, x, t_input, **model_kwargs)
224
+ ``
225
+ where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
226
+ ===============================================================
227
+ Args:
228
+ model: A diffusion model with the corresponding format described above.
229
+ noise_schedule: A noise schedule object, such as NoiseScheduleVP.
230
+ model_type: A `str`. The parameterization type of the diffusion model.
231
+ "noise" or "x_start" or "v" or "score".
232
+ model_kwargs: A `dict`. A dict for the other inputs of the model function.
233
+ guidance_type: A `str`. The type of the guidance for sampling.
234
+ "uncond" or "classifier" or "classifier-free".
235
+ condition: A pytorch tensor. The condition for the guided sampling.
236
+ Only used for "classifier" or "classifier-free" guidance type.
237
+ unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
238
+ Only used for "classifier-free" guidance type.
239
+ guidance_scale: A `float`. The scale for the guided sampling.
240
+ classifier_fn: A classifier function. Only used for the classifier guidance.
241
+ classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
242
+ Returns:
243
+ A noise prediction model that accepts the noised data and the continuous time as the inputs.
244
+ """
245
+
246
+ def get_model_input_time(t_continuous):
247
+ """
248
+ Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
249
+ For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
250
+ For continuous-time DPMs, we just use `t_continuous`.
251
+ """
252
+ if noise_schedule.schedule == 'discrete':
253
+ return (t_continuous - 1. / noise_schedule.total_N) * 1000.
254
+ else:
255
+ return t_continuous
256
+
257
+ def noise_pred_fn(x, t_continuous, cond=None):
258
+ if t_continuous.reshape((-1,)).shape[0] == 1:
259
+ t_continuous = t_continuous.expand((x.shape[0]))
260
+ t_input = get_model_input_time(t_continuous)
261
+ if cond is None:
262
+ output = model(x, t_input, **model_kwargs)
263
+ else:
264
+ output = model(x, t_input, cond, **model_kwargs)
265
+ if model_type == "noise":
266
+ return output
267
+ elif model_type == "x_start":
268
+ alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
269
+ dims = x.dim()
270
+ return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims)
271
+ elif model_type == "v":
272
+ alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
273
+ dims = x.dim()
274
+ return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x
275
+ elif model_type == "score":
276
+ sigma_t = noise_schedule.marginal_std(t_continuous)
277
+ dims = x.dim()
278
+ return -expand_dims(sigma_t, dims) * output
279
+
280
+ def cond_grad_fn(x, t_input):
281
+ """
282
+ Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
283
+ """
284
+ with torch.enable_grad():
285
+ x_in = x.detach().requires_grad_(True)
286
+ log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
287
+ return torch.autograd.grad(log_prob.sum(), x_in)[0]
288
+
289
+ def model_fn(x, t_continuous):
290
+ """
291
+ The noise predicition model function that is used for DPM-Solver.
292
+ """
293
+ if t_continuous.reshape((-1,)).shape[0] == 1:
294
+ t_continuous = t_continuous.expand((x.shape[0]))
295
+ if guidance_type == "uncond":
296
+ return noise_pred_fn(x, t_continuous)
297
+ elif guidance_type == "classifier":
298
+ assert classifier_fn is not None
299
+ t_input = get_model_input_time(t_continuous)
300
+ cond_grad = cond_grad_fn(x, t_input)
301
+ sigma_t = noise_schedule.marginal_std(t_continuous)
302
+ noise = noise_pred_fn(x, t_continuous)
303
+ return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad
304
+ elif guidance_type == "classifier-free":
305
+ if guidance_scale == 1. or unconditional_condition is None:
306
+ return noise_pred_fn(x, t_continuous, cond=condition)
307
+ else:
308
+ x_in = torch.cat([x] * 2)
309
+ t_in = torch.cat([t_continuous] * 2)
310
+ c_in = torch.cat([unconditional_condition, condition])
311
+ noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
312
+ return noise_uncond + guidance_scale * (noise - noise_uncond)
313
+
314
+ assert model_type in ["noise", "x_start", "v"]
315
+ assert guidance_type in ["uncond", "classifier", "classifier-free"]
316
+ return model_fn
317
+
318
+
319
+ class DPM_Solver:
320
+ def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.):
321
+ """Construct a DPM-Solver.
322
+ We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0").
323
+ If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver).
324
+ If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++).
325
+ In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True.
326
+ The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales.
327
+ Args:
328
+ model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
329
+ ``
330
+ def model_fn(x, t_continuous):
331
+ return noise
332
+ ``
333
+ noise_schedule: A noise schedule object, such as NoiseScheduleVP.
334
+ predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model.
335
+ thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1].
336
+ max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding.
337
+
338
+ [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
339
+ """
340
+ self.model = model_fn
341
+ self.noise_schedule = noise_schedule
342
+ self.predict_x0 = predict_x0
343
+ self.thresholding = thresholding
344
+ self.max_val = max_val
345
+
346
+ def noise_prediction_fn(self, x, t):
347
+ """
348
+ Return the noise prediction model.
349
+ """
350
+ return self.model(x, t)
351
+
352
+ def data_prediction_fn(self, x, t):
353
+ """
354
+ Return the data prediction model (with thresholding).
355
+ """
356
+ noise = self.noise_prediction_fn(x, t)
357
+ dims = x.dim()
358
+ alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
359
+ x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims)
360
+ if self.thresholding:
361
+ p = 0.995 # A hyperparameter in the paper of "Imagen" [1].
362
+ s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
363
+ s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
364
+ x0 = torch.clamp(x0, -s, s) / s
365
+ return x0
366
+
367
+ def model_fn(self, x, t):
368
+ """
369
+ Convert the model to the noise prediction model or the data prediction model.
370
+ """
371
+ if self.predict_x0:
372
+ return self.data_prediction_fn(x, t)
373
+ else:
374
+ return self.noise_prediction_fn(x, t)
375
+
376
+ def get_time_steps(self, skip_type, t_T, t_0, N, device):
377
+ """Compute the intermediate time steps for sampling.
378
+ Args:
379
+ skip_type: A `str`. The type for the spacing of the time steps. We support three types:
380
+ - 'logSNR': uniform logSNR for the time steps.
381
+ - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
382
+ - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
383
+ t_T: A `float`. The starting time of the sampling (default is T).
384
+ t_0: A `float`. The ending time of the sampling (default is epsilon).
385
+ N: A `int`. The total number of the spacing of the time steps.
386
+ device: A torch device.
387
+ Returns:
388
+ A pytorch tensor of the time steps, with the shape (N + 1,).
389
+ """
390
+ if skip_type == 'logSNR':
391
+ lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
392
+ lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
393
+ logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
394
+ return self.noise_schedule.inverse_lambda(logSNR_steps)
395
+ elif skip_type == 'time_uniform':
396
+ return torch.linspace(t_T, t_0, N + 1).to(device)
397
+ elif skip_type == 'time_quadratic':
398
+ t_order = 2
399
+ t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device)
400
+ return t
401
+ else:
402
+ raise ValueError(
403
+ "Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type))
404
+
405
+ def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
406
+ """
407
+ Get the order of each step for sampling by the singlestep DPM-Solver.
408
+ We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
409
+ Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
410
+ - If order == 1:
411
+ We take `steps` of DPM-Solver-1 (i.e. DDIM).
412
+ - If order == 2:
413
+ - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
414
+ - If steps % 2 == 0, we use K steps of DPM-Solver-2.
415
+ - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
416
+ - If order == 3:
417
+ - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
418
+ - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
419
+ - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
420
+ - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
421
+ ============================================
422
+ Args:
423
+ order: A `int`. The max order for the solver (2 or 3).
424
+ steps: A `int`. The total number of function evaluations (NFE).
425
+ skip_type: A `str`. The type for the spacing of the time steps. We support three types:
426
+ - 'logSNR': uniform logSNR for the time steps.
427
+ - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
428
+ - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
429
+ t_T: A `float`. The starting time of the sampling (default is T).
430
+ t_0: A `float`. The ending time of the sampling (default is epsilon).
431
+ device: A torch device.
432
+ Returns:
433
+ orders: A list of the solver order of each step.
434
+ """
435
+ if order == 3:
436
+ K = steps // 3 + 1
437
+ if steps % 3 == 0:
438
+ orders = [3, ] * (K - 2) + [2, 1]
439
+ elif steps % 3 == 1:
440
+ orders = [3, ] * (K - 1) + [1]
441
+ else:
442
+ orders = [3, ] * (K - 1) + [2]
443
+ elif order == 2:
444
+ if steps % 2 == 0:
445
+ K = steps // 2
446
+ orders = [2, ] * K
447
+ else:
448
+ K = steps // 2 + 1
449
+ orders = [2, ] * (K - 1) + [1]
450
+ elif order == 1:
451
+ K = 1
452
+ orders = [1, ] * steps
453
+ else:
454
+ raise ValueError("'order' must be '1' or '2' or '3'.")
455
+ if skip_type == 'logSNR':
456
+ # To reproduce the results in DPM-Solver paper
457
+ timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
458
+ else:
459
+ timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
460
+ torch.cumsum(torch.tensor([0, ] + orders)).to(device)]
461
+ return timesteps_outer, orders
462
+
463
+ def denoise_to_zero_fn(self, x, s):
464
+ """
465
+ Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
466
+ """
467
+ return self.data_prediction_fn(x, s)
468
+
469
+ def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
470
+ """
471
+ DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
472
+ Args:
473
+ x: A pytorch tensor. The initial value at time `s`.
474
+ s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
475
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
476
+ model_s: A pytorch tensor. The model function evaluated at time `s`.
477
+ If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
478
+ return_intermediate: A `bool`. If true, also return the model value at time `s`.
479
+ Returns:
480
+ x_t: A pytorch tensor. The approximated solution at time `t`.
481
+ """
482
+ ns = self.noise_schedule
483
+ dims = x.dim()
484
+ lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
485
+ h = lambda_t - lambda_s
486
+ log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t)
487
+ sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
488
+ alpha_t = torch.exp(log_alpha_t)
489
+
490
+ if self.predict_x0:
491
+ phi_1 = torch.expm1(-h)
492
+ if model_s is None:
493
+ model_s = self.model_fn(x, s)
494
+ x_t = (
495
+ expand_dims(sigma_t / sigma_s, dims) * x
496
+ - expand_dims(alpha_t * phi_1, dims) * model_s
497
+ )
498
+ if return_intermediate:
499
+ return x_t, {'model_s': model_s}
500
+ else:
501
+ return x_t
502
+ else:
503
+ phi_1 = torch.expm1(h)
504
+ if model_s is None:
505
+ model_s = self.model_fn(x, s)
506
+ x_t = (
507
+ expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
508
+ - expand_dims(sigma_t * phi_1, dims) * model_s
509
+ )
510
+ if return_intermediate:
511
+ return x_t, {'model_s': model_s}
512
+ else:
513
+ return x_t
514
+
515
+ def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False,
516
+ solver_type='dpm_solver'):
517
+ """
518
+ Singlestep solver DPM-Solver-2 from time `s` to time `t`.
519
+ Args:
520
+ x: A pytorch tensor. The initial value at time `s`.
521
+ s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
522
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
523
+ r1: A `float`. The hyperparameter of the second-order solver.
524
+ model_s: A pytorch tensor. The model function evaluated at time `s`.
525
+ If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
526
+ return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
527
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
528
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
529
+ Returns:
530
+ x_t: A pytorch tensor. The approximated solution at time `t`.
531
+ """
532
+ if solver_type not in ['dpm_solver', 'taylor']:
533
+ raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
534
+ if r1 is None:
535
+ r1 = 0.5
536
+ ns = self.noise_schedule
537
+ dims = x.dim()
538
+ lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
539
+ h = lambda_t - lambda_s
540
+ lambda_s1 = lambda_s + r1 * h
541
+ s1 = ns.inverse_lambda(lambda_s1)
542
+ log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(
543
+ s1), ns.marginal_log_mean_coeff(t)
544
+ sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t)
545
+ alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
546
+
547
+ if self.predict_x0:
548
+ phi_11 = torch.expm1(-r1 * h)
549
+ phi_1 = torch.expm1(-h)
550
+
551
+ if model_s is None:
552
+ model_s = self.model_fn(x, s)
553
+ x_s1 = (
554
+ expand_dims(sigma_s1 / sigma_s, dims) * x
555
+ - expand_dims(alpha_s1 * phi_11, dims) * model_s
556
+ )
557
+ model_s1 = self.model_fn(x_s1, s1)
558
+ if solver_type == 'dpm_solver':
559
+ x_t = (
560
+ expand_dims(sigma_t / sigma_s, dims) * x
561
+ - expand_dims(alpha_t * phi_1, dims) * model_s
562
+ - (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s)
563
+ )
564
+ elif solver_type == 'taylor':
565
+ x_t = (
566
+ expand_dims(sigma_t / sigma_s, dims) * x
567
+ - expand_dims(alpha_t * phi_1, dims) * model_s
568
+ + (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * (
569
+ model_s1 - model_s)
570
+ )
571
+ else:
572
+ phi_11 = torch.expm1(r1 * h)
573
+ phi_1 = torch.expm1(h)
574
+
575
+ if model_s is None:
576
+ model_s = self.model_fn(x, s)
577
+ x_s1 = (
578
+ expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
579
+ - expand_dims(sigma_s1 * phi_11, dims) * model_s
580
+ )
581
+ model_s1 = self.model_fn(x_s1, s1)
582
+ if solver_type == 'dpm_solver':
583
+ x_t = (
584
+ expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
585
+ - expand_dims(sigma_t * phi_1, dims) * model_s
586
+ - (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s)
587
+ )
588
+ elif solver_type == 'taylor':
589
+ x_t = (
590
+ expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
591
+ - expand_dims(sigma_t * phi_1, dims) * model_s
592
+ - (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s)
593
+ )
594
+ if return_intermediate:
595
+ return x_t, {'model_s': model_s, 'model_s1': model_s1}
596
+ else:
597
+ return x_t
598
+
599
+ def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None,
600
+ return_intermediate=False, solver_type='dpm_solver'):
601
+ """
602
+ Singlestep solver DPM-Solver-3 from time `s` to time `t`.
603
+ Args:
604
+ x: A pytorch tensor. The initial value at time `s`.
605
+ s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
606
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
607
+ r1: A `float`. The hyperparameter of the third-order solver.
608
+ r2: A `float`. The hyperparameter of the third-order solver.
609
+ model_s: A pytorch tensor. The model function evaluated at time `s`.
610
+ If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
611
+ model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
612
+ If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
613
+ return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
614
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
615
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
616
+ Returns:
617
+ x_t: A pytorch tensor. The approximated solution at time `t`.
618
+ """
619
+ if solver_type not in ['dpm_solver', 'taylor']:
620
+ raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
621
+ if r1 is None:
622
+ r1 = 1. / 3.
623
+ if r2 is None:
624
+ r2 = 2. / 3.
625
+ ns = self.noise_schedule
626
+ dims = x.dim()
627
+ lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
628
+ h = lambda_t - lambda_s
629
+ lambda_s1 = lambda_s + r1 * h
630
+ lambda_s2 = lambda_s + r2 * h
631
+ s1 = ns.inverse_lambda(lambda_s1)
632
+ s2 = ns.inverse_lambda(lambda_s2)
633
+ log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff(
634
+ s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t)
635
+ sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(
636
+ s2), ns.marginal_std(t)
637
+ alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t)
638
+
639
+ if self.predict_x0:
640
+ phi_11 = torch.expm1(-r1 * h)
641
+ phi_12 = torch.expm1(-r2 * h)
642
+ phi_1 = torch.expm1(-h)
643
+ phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.
644
+ phi_2 = phi_1 / h + 1.
645
+ phi_3 = phi_2 / h - 0.5
646
+
647
+ if model_s is None:
648
+ model_s = self.model_fn(x, s)
649
+ if model_s1 is None:
650
+ x_s1 = (
651
+ expand_dims(sigma_s1 / sigma_s, dims) * x
652
+ - expand_dims(alpha_s1 * phi_11, dims) * model_s
653
+ )
654
+ model_s1 = self.model_fn(x_s1, s1)
655
+ x_s2 = (
656
+ expand_dims(sigma_s2 / sigma_s, dims) * x
657
+ - expand_dims(alpha_s2 * phi_12, dims) * model_s
658
+ + r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s)
659
+ )
660
+ model_s2 = self.model_fn(x_s2, s2)
661
+ if solver_type == 'dpm_solver':
662
+ x_t = (
663
+ expand_dims(sigma_t / sigma_s, dims) * x
664
+ - expand_dims(alpha_t * phi_1, dims) * model_s
665
+ + (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s)
666
+ )
667
+ elif solver_type == 'taylor':
668
+ D1_0 = (1. / r1) * (model_s1 - model_s)
669
+ D1_1 = (1. / r2) * (model_s2 - model_s)
670
+ D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
671
+ D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
672
+ x_t = (
673
+ expand_dims(sigma_t / sigma_s, dims) * x
674
+ - expand_dims(alpha_t * phi_1, dims) * model_s
675
+ + expand_dims(alpha_t * phi_2, dims) * D1
676
+ - expand_dims(alpha_t * phi_3, dims) * D2
677
+ )
678
+ else:
679
+ phi_11 = torch.expm1(r1 * h)
680
+ phi_12 = torch.expm1(r2 * h)
681
+ phi_1 = torch.expm1(h)
682
+ phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.
683
+ phi_2 = phi_1 / h - 1.
684
+ phi_3 = phi_2 / h - 0.5
685
+
686
+ if model_s is None:
687
+ model_s = self.model_fn(x, s)
688
+ if model_s1 is None:
689
+ x_s1 = (
690
+ expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
691
+ - expand_dims(sigma_s1 * phi_11, dims) * model_s
692
+ )
693
+ model_s1 = self.model_fn(x_s1, s1)
694
+ x_s2 = (
695
+ expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x
696
+ - expand_dims(sigma_s2 * phi_12, dims) * model_s
697
+ - r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s)
698
+ )
699
+ model_s2 = self.model_fn(x_s2, s2)
700
+ if solver_type == 'dpm_solver':
701
+ x_t = (
702
+ expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
703
+ - expand_dims(sigma_t * phi_1, dims) * model_s
704
+ - (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s)
705
+ )
706
+ elif solver_type == 'taylor':
707
+ D1_0 = (1. / r1) * (model_s1 - model_s)
708
+ D1_1 = (1. / r2) * (model_s2 - model_s)
709
+ D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
710
+ D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
711
+ x_t = (
712
+ expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
713
+ - expand_dims(sigma_t * phi_1, dims) * model_s
714
+ - expand_dims(sigma_t * phi_2, dims) * D1
715
+ - expand_dims(sigma_t * phi_3, dims) * D2
716
+ )
717
+
718
+ if return_intermediate:
719
+ return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2}
720
+ else:
721
+ return x_t
722
+
723
+ def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"):
724
+ """
725
+ Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
726
+ Args:
727
+ x: A pytorch tensor. The initial value at time `s`.
728
+ model_prev_list: A list of pytorch tensor. The previous computed model values.
729
+ t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
730
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
731
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
732
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
733
+ Returns:
734
+ x_t: A pytorch tensor. The approximated solution at time `t`.
735
+ """
736
+ if solver_type not in ['dpm_solver', 'taylor']:
737
+ raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
738
+ ns = self.noise_schedule
739
+ dims = x.dim()
740
+ model_prev_1, model_prev_0 = model_prev_list
741
+ t_prev_1, t_prev_0 = t_prev_list
742
+ lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda(
743
+ t_prev_0), ns.marginal_lambda(t)
744
+ log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
745
+ sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
746
+ alpha_t = torch.exp(log_alpha_t)
747
+
748
+ h_0 = lambda_prev_0 - lambda_prev_1
749
+ h = lambda_t - lambda_prev_0
750
+ r0 = h_0 / h
751
+ D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
752
+ if self.predict_x0:
753
+ if solver_type == 'dpm_solver':
754
+ x_t = (
755
+ expand_dims(sigma_t / sigma_prev_0, dims) * x
756
+ - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
757
+ - 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0
758
+ )
759
+ elif solver_type == 'taylor':
760
+ x_t = (
761
+ expand_dims(sigma_t / sigma_prev_0, dims) * x
762
+ - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
763
+ + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0
764
+ )
765
+ else:
766
+ if solver_type == 'dpm_solver':
767
+ x_t = (
768
+ expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
769
+ - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
770
+ - 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0
771
+ )
772
+ elif solver_type == 'taylor':
773
+ x_t = (
774
+ expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
775
+ - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
776
+ - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0
777
+ )
778
+ return x_t
779
+
780
+ def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'):
781
+ """
782
+ Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
783
+ Args:
784
+ x: A pytorch tensor. The initial value at time `s`.
785
+ model_prev_list: A list of pytorch tensor. The previous computed model values.
786
+ t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
787
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
788
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
789
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
790
+ Returns:
791
+ x_t: A pytorch tensor. The approximated solution at time `t`.
792
+ """
793
+ ns = self.noise_schedule
794
+ dims = x.dim()
795
+ model_prev_2, model_prev_1, model_prev_0 = model_prev_list
796
+ t_prev_2, t_prev_1, t_prev_0 = t_prev_list
797
+ lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda(
798
+ t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
799
+ log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
800
+ sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
801
+ alpha_t = torch.exp(log_alpha_t)
802
+
803
+ h_1 = lambda_prev_1 - lambda_prev_2
804
+ h_0 = lambda_prev_0 - lambda_prev_1
805
+ h = lambda_t - lambda_prev_0
806
+ r0, r1 = h_0 / h, h_1 / h
807
+ D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
808
+ D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2)
809
+ D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1)
810
+ D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1)
811
+ if self.predict_x0:
812
+ x_t = (
813
+ expand_dims(sigma_t / sigma_prev_0, dims) * x
814
+ - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
815
+ + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1
816
+ - expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2
817
+ )
818
+ else:
819
+ x_t = (
820
+ expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
821
+ - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
822
+ - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1
823
+ - expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2
824
+ )
825
+ return x_t
826
+
827
+ def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None,
828
+ r2=None):
829
+ """
830
+ Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
831
+ Args:
832
+ x: A pytorch tensor. The initial value at time `s`.
833
+ s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
834
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
835
+ order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
836
+ return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
837
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
838
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
839
+ r1: A `float`. The hyperparameter of the second-order or third-order solver.
840
+ r2: A `float`. The hyperparameter of the third-order solver.
841
+ Returns:
842
+ x_t: A pytorch tensor. The approximated solution at time `t`.
843
+ """
844
+ if order == 1:
845
+ return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate)
846
+ elif order == 2:
847
+ return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate,
848
+ solver_type=solver_type, r1=r1)
849
+ elif order == 3:
850
+ return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate,
851
+ solver_type=solver_type, r1=r1, r2=r2)
852
+ else:
853
+ raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
854
+
855
+ def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'):
856
+ """
857
+ Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
858
+ Args:
859
+ x: A pytorch tensor. The initial value at time `s`.
860
+ model_prev_list: A list of pytorch tensor. The previous computed model values.
861
+ t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
862
+ t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
863
+ order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
864
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
865
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
866
+ Returns:
867
+ x_t: A pytorch tensor. The approximated solution at time `t`.
868
+ """
869
+ if order == 1:
870
+ return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1])
871
+ elif order == 2:
872
+ return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
873
+ elif order == 3:
874
+ return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
875
+ else:
876
+ raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
877
+
878
+ def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5,
879
+ solver_type='dpm_solver'):
880
+ """
881
+ The adaptive step size solver based on singlestep DPM-Solver.
882
+ Args:
883
+ x: A pytorch tensor. The initial value at time `t_T`.
884
+ order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
885
+ t_T: A `float`. The starting time of the sampling (default is T).
886
+ t_0: A `float`. The ending time of the sampling (default is epsilon).
887
+ h_init: A `float`. The initial step size (for logSNR).
888
+ atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
889
+ rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
890
+ theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
891
+ t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
892
+ current time and `t_0` is less than `t_err`. The default setting is 1e-5.
893
+ solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
894
+ The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
895
+ Returns:
896
+ x_0: A pytorch tensor. The approximated solution at time `t_0`.
897
+ [1] A. Jolicoeur-Martineau, K. Li, R. PichΓ©-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
898
+ """
899
+ ns = self.noise_schedule
900
+ s = t_T * torch.ones((x.shape[0],)).to(x)
901
+ lambda_s = ns.marginal_lambda(s)
902
+ lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
903
+ h = h_init * torch.ones_like(s).to(x)
904
+ x_prev = x
905
+ nfe = 0
906
+ if order == 2:
907
+ r1 = 0.5
908
+ lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True)
909
+ higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1,
910
+ solver_type=solver_type,
911
+ **kwargs)
912
+ elif order == 3:
913
+ r1, r2 = 1. / 3., 2. / 3.
914
+ lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1,
915
+ return_intermediate=True,
916
+ solver_type=solver_type)
917
+ higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2,
918
+ solver_type=solver_type,
919
+ **kwargs)
920
+ else:
921
+ raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order))
922
+ while torch.abs((s - t_0)).mean() > t_err:
923
+ t = ns.inverse_lambda(lambda_s + h)
924
+ x_lower, lower_noise_kwargs = lower_update(x, s, t)
925
+ x_higher = higher_update(x, s, t, **lower_noise_kwargs)
926
+ delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)))
927
+ norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True))
928
+ E = norm_fn((x_higher - x_lower) / delta).max()
929
+ if torch.all(E <= 1.):
930
+ x = x_higher
931
+ s = t
932
+ x_prev = x_lower
933
+ lambda_s = ns.marginal_lambda(s)
934
+ h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s)
935
+ nfe += order
936
+ print('adaptive solver nfe', nfe)
937
+ return x
938
+
939
+ def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform',
940
+ method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver',
941
+ atol=0.0078, rtol=0.05,
942
+ ):
943
+ """
944
+ Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
945
+ =====================================================
946
+ We support the following algorithms for both noise prediction model and data prediction model:
947
+ - 'singlestep':
948
+ Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
949
+ We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
950
+ The total number of function evaluations (NFE) == `steps`.
951
+ Given a fixed NFE == `steps`, the sampling procedure is:
952
+ - If `order` == 1:
953
+ - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
954
+ - If `order` == 2:
955
+ - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
956
+ - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
957
+ - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
958
+ - If `order` == 3:
959
+ - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
960
+ - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
961
+ - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
962
+ - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
963
+ - 'multistep':
964
+ Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
965
+ We initialize the first `order` values by lower order multistep solvers.
966
+ Given a fixed NFE == `steps`, the sampling procedure is:
967
+ Denote K = steps.
968
+ - If `order` == 1:
969
+ - We use K steps of DPM-Solver-1 (i.e. DDIM).
970
+ - If `order` == 2:
971
+ - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
972
+ - If `order` == 3:
973
+ - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
974
+ - 'singlestep_fixed':
975
+ Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
976
+ We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
977
+ - 'adaptive':
978
+ Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
979
+ We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
980
+ You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
981
+ (NFE) and the sample quality.
982
+ - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
983
+ - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
984
+ =====================================================
985
+ Some advices for choosing the algorithm:
986
+ - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
987
+ Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`.
988
+ e.g.
989
+ >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False)
990
+ >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
991
+ skip_type='time_uniform', method='singlestep')
992
+ - For **guided sampling with large guidance scale** by DPMs:
993
+ Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`.
994
+ e.g.
995
+ >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True)
996
+ >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
997
+ skip_type='time_uniform', method='multistep')
998
+ We support three types of `skip_type`:
999
+ - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
1000
+ - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
1001
+ - 'time_quadratic': quadratic time for the time steps.
1002
+ =====================================================
1003
+ Args:
1004
+ x: A pytorch tensor. The initial value at time `t_start`
1005
+ e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
1006
+ steps: A `int`. The total number of function evaluations (NFE).
1007
+ t_start: A `float`. The starting time of the sampling.
1008
+ If `T` is None, we use self.noise_schedule.T (default is 1.0).
1009
+ t_end: A `float`. The ending time of the sampling.
1010
+ If `t_end` is None, we use 1. / self.noise_schedule.total_N.
1011
+ e.g. if total_N == 1000, we have `t_end` == 1e-3.
1012
+ For discrete-time DPMs:
1013
+ - We recommend `t_end` == 1. / self.noise_schedule.total_N.
1014
+ For continuous-time DPMs:
1015
+ - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
1016
+ order: A `int`. The order of DPM-Solver.
1017
+ skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
1018
+ method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
1019
+ denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
1020
+ Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
1021
+ This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
1022
+ score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
1023
+ for diffusion models sampling by diffusion SDEs for low-resolutional images
1024
+ (such as CIFAR-10). However, we observed that such trick does not matter for
1025
+ high-resolutional images. As it needs an additional NFE, we do not recommend
1026
+ it for high-resolutional images.
1027
+ lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
1028
+ Only valid for `method=multistep` and `steps < 15`. We empirically find that
1029
+ this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
1030
+ (especially for steps <= 10). So we recommend to set it to be `True`.
1031
+ solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`.
1032
+ atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
1033
+ rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
1034
+ Returns:
1035
+ x_end: A pytorch tensor. The approximated solution at time `t_end`.
1036
+ """
1037
+ t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end
1038
+ t_T = self.noise_schedule.T if t_start is None else t_start
1039
+ device = x.device
1040
+ if method == 'adaptive':
1041
+ with torch.no_grad():
1042
+ x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol,
1043
+ solver_type=solver_type)
1044
+ elif method == 'multistep':
1045
+ assert steps >= order
1046
+ timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device)
1047
+ assert timesteps.shape[0] - 1 == steps
1048
+ with torch.no_grad():
1049
+ vec_t = timesteps[0].expand((x.shape[0]))
1050
+ model_prev_list = [self.model_fn(x, vec_t)]
1051
+ t_prev_list = [vec_t]
1052
+ # Init the first `order` values by lower order multistep DPM-Solver.
1053
+ for init_order in tqdm(range(1, order), desc="DPM init order"):
1054
+ vec_t = timesteps[init_order].expand(x.shape[0])
1055
+ x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order,
1056
+ solver_type=solver_type)
1057
+ model_prev_list.append(self.model_fn(x, vec_t))
1058
+ t_prev_list.append(vec_t)
1059
+ # Compute the remaining values by `order`-th order multistep DPM-Solver.
1060
+ for step in tqdm(range(order, steps + 1), desc="DPM multistep"):
1061
+ vec_t = timesteps[step].expand(x.shape[0])
1062
+ if lower_order_final and steps < 15:
1063
+ step_order = min(order, steps + 1 - step)
1064
+ else:
1065
+ step_order = order
1066
+ x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order,
1067
+ solver_type=solver_type)
1068
+ for i in range(order - 1):
1069
+ t_prev_list[i] = t_prev_list[i + 1]
1070
+ model_prev_list[i] = model_prev_list[i + 1]
1071
+ t_prev_list[-1] = vec_t
1072
+ # We do not need to evaluate the final model value.
1073
+ if step < steps:
1074
+ model_prev_list[-1] = self.model_fn(x, vec_t)
1075
+ elif method in ['singlestep', 'singlestep_fixed']:
1076
+ if method == 'singlestep':
1077
+ timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order,
1078
+ skip_type=skip_type,
1079
+ t_T=t_T, t_0=t_0,
1080
+ device=device)
1081
+ elif method == 'singlestep_fixed':
1082
+ K = steps // order
1083
+ orders = [order, ] * K
1084
+ timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device)
1085
+ for i, order in enumerate(orders):
1086
+ t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1]
1087
+ timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(),
1088
+ N=order, device=device)
1089
+ lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
1090
+ vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0])
1091
+ h = lambda_inner[-1] - lambda_inner[0]
1092
+ r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
1093
+ r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
1094
+ x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2)
1095
+ if denoise_to_zero:
1096
+ x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0)
1097
+ return x
1098
+
1099
+
1100
+ #############################################################
1101
+ # other utility functions
1102
+ #############################################################
1103
+
1104
+ def interpolate_fn(x, xp, yp):
1105
+ """
1106
+ A piecewise linear function y = f(x), using xp and yp as keypoints.
1107
+ We implement f(x) in a differentiable way (i.e. applicable for autograd).
1108
+ The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
1109
+ Args:
1110
+ x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
1111
+ xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
1112
+ yp: PyTorch tensor with shape [C, K].
1113
+ Returns:
1114
+ The function values f(x), with shape [N, C].
1115
+ """
1116
+ N, K = x.shape[0], xp.shape[1]
1117
+ all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
1118
+ sorted_all_x, x_indices = torch.sort(all_x, dim=2)
1119
+ x_idx = torch.argmin(x_indices, dim=2)
1120
+ cand_start_idx = x_idx - 1
1121
+ start_idx = torch.where(
1122
+ torch.eq(x_idx, 0),
1123
+ torch.tensor(1, device=x.device),
1124
+ torch.where(
1125
+ torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
1126
+ ),
1127
+ )
1128
+ end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
1129
+ start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
1130
+ end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
1131
+ start_idx2 = torch.where(
1132
+ torch.eq(x_idx, 0),
1133
+ torch.tensor(0, device=x.device),
1134
+ torch.where(
1135
+ torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
1136
+ ),
1137
+ )
1138
+ y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
1139
+ start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
1140
+ end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
1141
+ cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
1142
+ return cand
1143
+
1144
+
1145
+ def expand_dims(v, dims):
1146
+ """
1147
+ Expand the tensor `v` to the dim `dims`.
1148
+ Args:
1149
+ `v`: a PyTorch tensor with shape [N].
1150
+ `dim`: a `int`.
1151
+ Returns:
1152
+ a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
1153
+ """
1154
+ return v[(...,) + (None,) * (dims - 1)]
ldm/models/diffusion/dpm_solver/sampler.py ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """SAMPLING ONLY."""
2
+ import torch
3
+
4
+ from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver
5
+
6
+
7
+ MODEL_TYPES = {
8
+ "eps": "noise",
9
+ "v": "v"
10
+ }
11
+
12
+
13
+ class DPMSolverSampler(object):
14
+ def __init__(self, model, **kwargs):
15
+ super().__init__()
16
+ self.model = model
17
+ to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device)
18
+ self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod))
19
+
20
+ def register_buffer(self, name, attr):
21
+ if type(attr) == torch.Tensor:
22
+ if attr.device != torch.device("cuda"):
23
+ attr = attr.to(torch.device("cuda"))
24
+ setattr(self, name, attr)
25
+
26
+ @torch.no_grad()
27
+ def sample(self,
28
+ S,
29
+ batch_size,
30
+ shape,
31
+ conditioning=None,
32
+ callback=None,
33
+ normals_sequence=None,
34
+ img_callback=None,
35
+ quantize_x0=False,
36
+ eta=0.,
37
+ mask=None,
38
+ x0=None,
39
+ temperature=1.,
40
+ noise_dropout=0.,
41
+ score_corrector=None,
42
+ corrector_kwargs=None,
43
+ verbose=True,
44
+ x_T=None,
45
+ log_every_t=100,
46
+ unconditional_guidance_scale=1.,
47
+ unconditional_conditioning=None,
48
+ # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
49
+ **kwargs
50
+ ):
51
+ if conditioning is not None:
52
+ if isinstance(conditioning, dict):
53
+ cbs = conditioning[list(conditioning.keys())[0]].shape[0]
54
+ if cbs != batch_size:
55
+ print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
56
+ else:
57
+ if conditioning.shape[0] != batch_size:
58
+ print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
59
+
60
+ # sampling
61
+ C, H, W = shape
62
+ size = (batch_size, C, H, W)
63
+
64
+ print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}')
65
+
66
+ device = self.model.betas.device
67
+ if x_T is None:
68
+ img = torch.randn(size, device=device)
69
+ else:
70
+ img = x_T
71
+
72
+ ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod)
73
+
74
+ model_fn = model_wrapper(
75
+ lambda x, t, c: self.model.apply_model(x, t, c),
76
+ ns,
77
+ model_type=MODEL_TYPES[self.model.parameterization],
78
+ guidance_type="classifier-free",
79
+ condition=conditioning,
80
+ unconditional_condition=unconditional_conditioning,
81
+ guidance_scale=unconditional_guidance_scale,
82
+ )
83
+
84
+ dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False)
85
+ x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True)
86
+
87
+ return x.to(device), None
ldm/models/diffusion/plms.py ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """SAMPLING ONLY."""
2
+
3
+ import torch
4
+ import numpy as np
5
+ from tqdm import tqdm
6
+ from functools import partial
7
+
8
+ from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
9
+ from ldm.models.diffusion.sampling_util import norm_thresholding
10
+
11
+
12
+ class PLMSSampler(object):
13
+ def __init__(self, model, schedule="linear", **kwargs):
14
+ super().__init__()
15
+ self.model = model
16
+ self.ddpm_num_timesteps = model.num_timesteps
17
+ self.schedule = schedule
18
+
19
+ def register_buffer(self, name, attr):
20
+ if type(attr) == torch.Tensor:
21
+ if attr.device != torch.device("cuda"):
22
+ attr = attr.to(torch.device("cuda"))
23
+ setattr(self, name, attr)
24
+
25
+ def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
26
+ if ddim_eta != 0:
27
+ raise ValueError('ddim_eta must be 0 for PLMS')
28
+ self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
29
+ num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
30
+ alphas_cumprod = self.model.alphas_cumprod
31
+ assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
32
+ to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
33
+
34
+ self.register_buffer('betas', to_torch(self.model.betas))
35
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
36
+ self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
37
+
38
+ # calculations for diffusion q(x_t | x_{t-1}) and others
39
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
40
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
41
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
42
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
43
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
44
+
45
+ # ddim sampling parameters
46
+ ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
47
+ ddim_timesteps=self.ddim_timesteps,
48
+ eta=ddim_eta,verbose=verbose)
49
+ self.register_buffer('ddim_sigmas', ddim_sigmas)
50
+ self.register_buffer('ddim_alphas', ddim_alphas)
51
+ self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
52
+ self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
53
+ sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
54
+ (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
55
+ 1 - self.alphas_cumprod / self.alphas_cumprod_prev))
56
+ self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
57
+
58
+ @torch.no_grad()
59
+ def sample(self,
60
+ S,
61
+ batch_size,
62
+ shape,
63
+ conditioning=None,
64
+ callback=None,
65
+ normals_sequence=None,
66
+ img_callback=None,
67
+ quantize_x0=False,
68
+ eta=0.,
69
+ mask=None,
70
+ x0=None,
71
+ temperature=1.,
72
+ noise_dropout=0.,
73
+ score_corrector=None,
74
+ corrector_kwargs=None,
75
+ verbose=True,
76
+ x_T=None,
77
+ log_every_t=100,
78
+ unconditional_guidance_scale=1.,
79
+ unconditional_conditioning=None,
80
+ # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
81
+ dynamic_threshold=None,
82
+ **kwargs
83
+ ):
84
+ if conditioning is not None:
85
+ if isinstance(conditioning, dict):
86
+ cbs = conditioning[list(conditioning.keys())[0]].shape[0]
87
+ if cbs != batch_size:
88
+ print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
89
+ else:
90
+ if conditioning.shape[0] != batch_size:
91
+ print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
92
+
93
+ self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
94
+ # sampling
95
+ C, H, W = shape
96
+ size = (batch_size, C, H, W)
97
+ print(f'Data shape for PLMS sampling is {size}')
98
+
99
+ samples, intermediates = self.plms_sampling(conditioning, size,
100
+ callback=callback,
101
+ img_callback=img_callback,
102
+ quantize_denoised=quantize_x0,
103
+ mask=mask, x0=x0,
104
+ ddim_use_original_steps=False,
105
+ noise_dropout=noise_dropout,
106
+ temperature=temperature,
107
+ score_corrector=score_corrector,
108
+ corrector_kwargs=corrector_kwargs,
109
+ x_T=x_T,
110
+ log_every_t=log_every_t,
111
+ unconditional_guidance_scale=unconditional_guidance_scale,
112
+ unconditional_conditioning=unconditional_conditioning,
113
+ dynamic_threshold=dynamic_threshold,
114
+ )
115
+ return samples, intermediates
116
+
117
+ @torch.no_grad()
118
+ def plms_sampling(self, cond, shape,
119
+ x_T=None, ddim_use_original_steps=False,
120
+ callback=None, timesteps=None, quantize_denoised=False,
121
+ mask=None, x0=None, img_callback=None, log_every_t=100,
122
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
123
+ unconditional_guidance_scale=1., unconditional_conditioning=None,
124
+ dynamic_threshold=None):
125
+ device = self.model.betas.device
126
+ b = shape[0]
127
+ if x_T is None:
128
+ img = torch.randn(shape, device=device)
129
+ else:
130
+ img = x_T
131
+
132
+ if timesteps is None:
133
+ timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
134
+ elif timesteps is not None and not ddim_use_original_steps:
135
+ subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
136
+ timesteps = self.ddim_timesteps[:subset_end]
137
+
138
+ intermediates = {'x_inter': [img], 'pred_x0': [img]}
139
+ time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
140
+ total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
141
+ print(f"Running PLMS Sampling with {total_steps} timesteps")
142
+
143
+ iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
144
+ old_eps = []
145
+
146
+ for i, step in enumerate(iterator):
147
+ index = total_steps - i - 1
148
+ ts = torch.full((b,), step, device=device, dtype=torch.long)
149
+ ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
150
+
151
+ if mask is not None:
152
+ assert x0 is not None
153
+ img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
154
+ img = img_orig * mask + (1. - mask) * img
155
+
156
+ outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
157
+ quantize_denoised=quantize_denoised, temperature=temperature,
158
+ noise_dropout=noise_dropout, score_corrector=score_corrector,
159
+ corrector_kwargs=corrector_kwargs,
160
+ unconditional_guidance_scale=unconditional_guidance_scale,
161
+ unconditional_conditioning=unconditional_conditioning,
162
+ old_eps=old_eps, t_next=ts_next,
163
+ dynamic_threshold=dynamic_threshold)
164
+ img, pred_x0, e_t = outs
165
+ old_eps.append(e_t)
166
+ if len(old_eps) >= 4:
167
+ old_eps.pop(0)
168
+ if callback: callback(i)
169
+ if img_callback: img_callback(pred_x0, i)
170
+
171
+ if index % log_every_t == 0 or index == total_steps - 1:
172
+ intermediates['x_inter'].append(img)
173
+ intermediates['pred_x0'].append(pred_x0)
174
+
175
+ return img, intermediates
176
+
177
+ @torch.no_grad()
178
+ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
179
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
180
+ unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,
181
+ dynamic_threshold=None):
182
+ b, *_, device = *x.shape, x.device
183
+
184
+ def get_model_output(x, t):
185
+ if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
186
+ e_t = self.model.apply_model(x, t, c)
187
+ else:
188
+ x_in = torch.cat([x] * 2)
189
+ t_in = torch.cat([t] * 2)
190
+ c_in = torch.cat([unconditional_conditioning, c])
191
+ e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
192
+ e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
193
+
194
+ if score_corrector is not None:
195
+ assert self.model.parameterization == "eps"
196
+ e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
197
+
198
+ return e_t
199
+
200
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
201
+ alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
202
+ sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
203
+ sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
204
+
205
+ def get_x_prev_and_pred_x0(e_t, index):
206
+ # select parameters corresponding to the currently considered timestep
207
+ a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
208
+ a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
209
+ sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
210
+ sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
211
+
212
+ # current prediction for x_0
213
+ pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
214
+ if quantize_denoised:
215
+ pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
216
+ if dynamic_threshold is not None:
217
+ pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
218
+ # direction pointing to x_t
219
+ dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
220
+ noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
221
+ if noise_dropout > 0.:
222
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
223
+ x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
224
+ return x_prev, pred_x0
225
+
226
+ e_t = get_model_output(x, t)
227
+ if len(old_eps) == 0:
228
+ # Pseudo Improved Euler (2nd order)
229
+ x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
230
+ e_t_next = get_model_output(x_prev, t_next)
231
+ e_t_prime = (e_t + e_t_next) / 2
232
+ elif len(old_eps) == 1:
233
+ # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
234
+ e_t_prime = (3 * e_t - old_eps[-1]) / 2
235
+ elif len(old_eps) == 2:
236
+ # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
237
+ e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
238
+ elif len(old_eps) >= 3:
239
+ # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
240
+ e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
241
+
242
+ x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
243
+
244
+ return x_prev, pred_x0, e_t
ldm/models/diffusion/sampling_util.py ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import numpy as np
3
+
4
+
5
+ def append_dims(x, target_dims):
6
+ """Appends dimensions to the end of a tensor until it has target_dims dimensions.
7
+ From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py"""
8
+ dims_to_append = target_dims - x.ndim
9
+ if dims_to_append < 0:
10
+ raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less')
11
+ return x[(...,) + (None,) * dims_to_append]
12
+
13
+
14
+ def norm_thresholding(x0, value):
15
+ s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim)
16
+ return x0 * (value / s)
17
+
18
+
19
+ def spatial_norm_thresholding(x0, value):
20
+ # b c h w
21
+ s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value)
22
+ return x0 * (value / s)
ldm/modules/attention.py ADDED
@@ -0,0 +1,330 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from inspect import isfunction
2
+ import math
3
+ import torch
4
+ import torch.nn.functional as F
5
+ from torch import nn, einsum
6
+ from einops import rearrange, repeat
7
+ from typing import Optional, Any
8
+ import os
9
+
10
+ from ldm.modules.diffusionmodules.util import checkpoint
11
+
12
+ try:
13
+ import xformers
14
+ import xformers.ops
15
+ XFORMERS_IS_AVAILBLE = True
16
+ except:
17
+ XFORMERS_IS_AVAILBLE = False
18
+
19
+ # CrossAttn precision handling
20
+ import os
21
+ _ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32")
22
+
23
+ def exists(val):
24
+ return val is not None
25
+
26
+
27
+ def uniq(arr):
28
+ return{el: True for el in arr}.keys()
29
+
30
+
31
+ def default(val, d):
32
+ if exists(val):
33
+ return val
34
+ return d() if isfunction(d) else d
35
+
36
+ class GEGLU(nn.Module):
37
+ def __init__(self, dim_in, dim_out):
38
+ super().__init__()
39
+ self.proj = nn.Linear(dim_in, dim_out * 2)
40
+
41
+ def forward(self, x):
42
+ x, gate = self.proj(x).chunk(2, dim=-1)
43
+ return x * F.gelu(gate)
44
+
45
+
46
+ class FeedForward(nn.Module):
47
+ def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
48
+ super().__init__()
49
+ inner_dim = int(dim * mult)
50
+ dim_out = default(dim_out, dim)
51
+ project_in = nn.Sequential(
52
+ nn.Linear(dim, inner_dim),
53
+ nn.GELU()
54
+ ) if not glu else GEGLU(dim, inner_dim)
55
+
56
+ self.net = nn.Sequential(
57
+ project_in,
58
+ nn.Dropout(dropout),
59
+ nn.Linear(inner_dim, dim_out)
60
+ )
61
+
62
+ def forward(self, x):
63
+ return self.net(x)
64
+
65
+
66
+ def zero_module(module):
67
+ """
68
+ Zero out the parameters of a module and return it.
69
+ """
70
+ for p in module.parameters():
71
+ p.detach().zero_()
72
+ return module
73
+
74
+
75
+ def Normalize(in_channels):
76
+ return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
77
+
78
+
79
+ class SpatialSelfAttention(nn.Module):
80
+ def __init__(self, in_channels):
81
+ super().__init__()
82
+ self.in_channels = in_channels
83
+
84
+ self.norm = Normalize(in_channels)
85
+ self.q = torch.nn.Conv2d(in_channels,
86
+ in_channels,
87
+ kernel_size=1,
88
+ stride=1,
89
+ padding=0)
90
+ self.k = torch.nn.Conv2d(in_channels,
91
+ in_channels,
92
+ kernel_size=1,
93
+ stride=1,
94
+ padding=0)
95
+ self.v = torch.nn.Conv2d(in_channels,
96
+ in_channels,
97
+ kernel_size=1,
98
+ stride=1,
99
+ padding=0)
100
+ self.proj_out = torch.nn.Conv2d(in_channels,
101
+ in_channels,
102
+ kernel_size=1,
103
+ stride=1,
104
+ padding=0)
105
+
106
+ def forward(self, x):
107
+ h_ = x
108
+ h_ = self.norm(h_)
109
+ q = self.q(h_)
110
+ k = self.k(h_)
111
+ v = self.v(h_)
112
+
113
+ b,c,h,w = q.shape
114
+ q = rearrange(q, 'b c h w -> b (h w) c')
115
+ k = rearrange(k, 'b c h w -> b c (h w)')
116
+ w_ = torch.einsum('bij,bjk->bik', q, k)
117
+
118
+ w_ = w_ * (int(c)**(-0.5))
119
+ w_ = torch.nn.functional.softmax(w_, dim=2)
120
+
121
+ v = rearrange(v, 'b c h w -> b c (h w)')
122
+ w_ = rearrange(w_, 'b i j -> b j i')
123
+ h_ = torch.einsum('bij,bjk->bik', v, w_)
124
+ h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
125
+ h_ = self.proj_out(h_)
126
+
127
+ return x+h_
128
+
129
+ class CrossAttention(nn.Module):
130
+ def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., **kwargs):
131
+ super().__init__()
132
+ inner_dim = dim_head * heads
133
+ context_dim = default(context_dim, query_dim)
134
+
135
+ self.scale = dim_head ** -0.5
136
+ self.heads = heads
137
+
138
+ self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
139
+ self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
140
+ self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
141
+
142
+ self.to_out = nn.Sequential(
143
+ nn.Linear(inner_dim, query_dim),
144
+ nn.Dropout(dropout)
145
+ )
146
+
147
+
148
+ def forward(self, x, context=None, mask=None):
149
+ h = self.heads
150
+ q = self.to_q(x)
151
+ context = default(context, x)
152
+ k = self.to_k(context)
153
+ v = self.to_v(context)
154
+ q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
155
+
156
+ if _ATTN_PRECISION =="fp32":
157
+ with torch.autocast(enabled=False, device_type = 'cuda'):
158
+ q, k = q.float(), k.float()
159
+ sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
160
+ else:
161
+ sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
162
+
163
+ del q, k
164
+ if exists(mask):
165
+ mask = rearrange(mask, 'b ... -> b (...)')
166
+ max_neg_value = -torch.finfo(sim.dtype).max
167
+ mask = repeat(mask, 'b j -> (b h) () j', h=h)
168
+ sim.masked_fill_(~mask, max_neg_value)
169
+
170
+ sim = sim.softmax(dim=-1)
171
+
172
+ out = einsum('b i j, b j d -> b i d', sim, v)
173
+ out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
174
+ return self.to_out(out)
175
+
176
+ class MemoryEfficientCrossAttention(nn.Module):
177
+ # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
178
+ def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, zero_init=False, **kwargs):
179
+ super().__init__()
180
+ print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
181
+ f"{heads} heads.")
182
+ inner_dim = dim_head * heads
183
+ context_dim = default(context_dim, query_dim)
184
+
185
+ self.heads = heads
186
+ self.dim_head = dim_head
187
+ if not zero_init:
188
+ self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
189
+ self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
190
+ self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
191
+ else:
192
+ self.to_q = zero_module(nn.Linear(query_dim, inner_dim, bias=False))
193
+ self.to_k = zero_module(nn.Linear(context_dim, inner_dim, bias=False))
194
+ self.to_v = zero_module(nn.Linear(context_dim, inner_dim, bias=False))
195
+
196
+ self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
197
+ self.attention_op: Optional[Any] = None
198
+
199
+
200
+ def forward(self, x, context=None, mask=None, **kwargs):
201
+ q = self.to_q(x)
202
+ context = default(context, x)
203
+ k = self.to_k(context)
204
+ v = self.to_v(context)
205
+ b, _, _ = q.shape
206
+ q, k, v = map(
207
+ lambda t: t.unsqueeze(3)
208
+ .reshape(b, t.shape[1], self.heads, self.dim_head)
209
+ .permute(0, 2, 1, 3)
210
+ .reshape(b * self.heads, t.shape[1], self.dim_head)
211
+ .contiguous(),
212
+ (q, k, v),
213
+ )
214
+ out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
215
+ if exists(mask):
216
+ raise NotImplementedError
217
+ out = (
218
+ out.unsqueeze(0)
219
+ .reshape(b, self.heads, out.shape[1], self.dim_head)
220
+ .permute(0, 2, 1, 3)
221
+ .reshape(b, out.shape[1], self.heads * self.dim_head)
222
+ )
223
+ return self.to_out(out)
224
+
225
+ class BasicTransformerBlock(nn.Module):
226
+ ATTENTION_MODES = {
227
+ "softmax": CrossAttention, # vanilla attention
228
+ "softmax-xformers": MemoryEfficientCrossAttention
229
+ }
230
+ def __init__(
231
+ self,
232
+ dim,
233
+ n_heads,
234
+ d_head,
235
+ dropout=0.,
236
+ context_dim=None,
237
+ gated_ff=True,
238
+ checkpoint=True,
239
+ disable_self_attn=False
240
+ ):
241
+ super().__init__()
242
+ attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax"
243
+ assert attn_mode in self.ATTENTION_MODES
244
+ attn_cls = self.ATTENTION_MODES[attn_mode]
245
+ self.disable_self_attn = disable_self_attn
246
+ self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
247
+ context_dim=context_dim if self.disable_self_attn else None)
248
+ self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
249
+ self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim,
250
+ heads=n_heads, dim_head=d_head, dropout=dropout)
251
+ self.norm1 = nn.LayerNorm(dim)
252
+ self.norm2 = nn.LayerNorm(dim)
253
+ self.norm3 = nn.LayerNorm(dim)
254
+ self.checkpoint = checkpoint
255
+
256
+ def forward(self, x, context=None,hint=None):
257
+ if hint is None:
258
+ return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
259
+ else:
260
+ return checkpoint(self._forward, (x, context, hint), self.parameters(), self.checkpoint)
261
+
262
+ def _forward(self, x, context=None,hint=None):
263
+ x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None,hint=hint) + x
264
+ x = self.attn2(self.norm2(x), context=context) + x
265
+ x = self.ff(self.norm3(x)) + x
266
+ return x
267
+
268
+ class SpatialTransformer(nn.Module):
269
+ """
270
+ Transformer block for image-like data.
271
+ First, project the input (aka embedding)
272
+ and reshape to b, t, d.
273
+ Then apply standard transformer action.
274
+ Finally, reshape to image
275
+ NEW: use_linear for more efficiency instead of the 1x1 convs
276
+ """
277
+ def __init__(self, in_channels, n_heads, d_head,
278
+ depth=1, dropout=0., context_dim=None,
279
+ disable_self_attn=False, use_linear=False,
280
+ use_checkpoint=True):
281
+ super().__init__()
282
+ if exists(context_dim) and not isinstance(context_dim, list):
283
+ context_dim = [context_dim]
284
+ self.in_channels = in_channels
285
+ inner_dim = n_heads * d_head
286
+ self.norm = Normalize(in_channels)
287
+ if not use_linear:
288
+ self.proj_in = nn.Conv2d(in_channels,
289
+ inner_dim,
290
+ kernel_size=1,
291
+ stride=1,
292
+ padding=0)
293
+ else:
294
+ self.proj_in = nn.Linear(in_channels, inner_dim)
295
+
296
+ self.transformer_blocks = nn.ModuleList(
297
+ [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
298
+ disable_self_attn=disable_self_attn, checkpoint=use_checkpoint)
299
+ for d in range(depth)]
300
+ )
301
+ if not use_linear:
302
+ self.proj_out = zero_module(nn.Conv2d(inner_dim,
303
+ in_channels,
304
+ kernel_size=1,
305
+ stride=1,
306
+ padding=0))
307
+ else:
308
+ self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
309
+ self.use_linear = use_linear
310
+
311
+ def forward(self, x, context=None,hint=None):
312
+ # note: if no context is given, cross-attention defaults to self-attention
313
+ if not isinstance(context, list):
314
+ context = [context]
315
+ b, c, h, w = x.shape
316
+ x_in = x
317
+ x = self.norm(x)
318
+ if not self.use_linear:
319
+ x = self.proj_in(x)
320
+ x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
321
+ if self.use_linear:
322
+ x = self.proj_in(x)
323
+ for i, block in enumerate(self.transformer_blocks):
324
+ x = block(x, context=context[i],hint=hint)
325
+ if self.use_linear:
326
+ x = self.proj_out(x)
327
+ x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
328
+ if not self.use_linear:
329
+ x = self.proj_out(x)
330
+ return x + x_in
ldm/modules/diffusionmodules/__init__.py ADDED
File without changes
ldm/modules/diffusionmodules/model.py ADDED
@@ -0,0 +1,852 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pytorch_diffusion + derived encoder decoder
2
+ import math
3
+ import torch
4
+ import torch.nn as nn
5
+ import numpy as np
6
+ from einops import rearrange
7
+ from typing import Optional, Any
8
+
9
+ from ldm.modules.attention import MemoryEfficientCrossAttention
10
+
11
+ try:
12
+ import xformers
13
+ import xformers.ops
14
+ XFORMERS_IS_AVAILBLE = True
15
+ except:
16
+ XFORMERS_IS_AVAILBLE = False
17
+ print("No module 'xformers'. Proceeding without it.")
18
+
19
+
20
+ def get_timestep_embedding(timesteps, embedding_dim):
21
+ """
22
+ This matches the implementation in Denoising Diffusion Probabilistic Models:
23
+ From Fairseq.
24
+ Build sinusoidal embeddings.
25
+ This matches the implementation in tensor2tensor, but differs slightly
26
+ from the description in Section 3.5 of "Attention Is All You Need".
27
+ """
28
+ assert len(timesteps.shape) == 1
29
+
30
+ half_dim = embedding_dim // 2
31
+ emb = math.log(10000) / (half_dim - 1)
32
+ emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
33
+ emb = emb.to(device=timesteps.device)
34
+ emb = timesteps.float()[:, None] * emb[None, :]
35
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
36
+ if embedding_dim % 2 == 1: # zero pad
37
+ emb = torch.nn.functional.pad(emb, (0,1,0,0))
38
+ return emb
39
+
40
+
41
+ def nonlinearity(x):
42
+ # swish
43
+ return x*torch.sigmoid(x)
44
+
45
+
46
+ def Normalize(in_channels, num_groups=32):
47
+ return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
48
+
49
+
50
+ class Upsample(nn.Module):
51
+ def __init__(self, in_channels, with_conv):
52
+ super().__init__()
53
+ self.with_conv = with_conv
54
+ if self.with_conv:
55
+ self.conv = torch.nn.Conv2d(in_channels,
56
+ in_channels,
57
+ kernel_size=3,
58
+ stride=1,
59
+ padding=1)
60
+
61
+ def forward(self, x):
62
+ x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
63
+ if self.with_conv:
64
+ x = self.conv(x)
65
+ return x
66
+
67
+
68
+ class Downsample(nn.Module):
69
+ def __init__(self, in_channels, with_conv):
70
+ super().__init__()
71
+ self.with_conv = with_conv
72
+ if self.with_conv:
73
+ # no asymmetric padding in torch conv, must do it ourselves
74
+ self.conv = torch.nn.Conv2d(in_channels,
75
+ in_channels,
76
+ kernel_size=3,
77
+ stride=2,
78
+ padding=0)
79
+
80
+ def forward(self, x):
81
+ if self.with_conv:
82
+ pad = (0,1,0,1)
83
+ x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
84
+ x = self.conv(x)
85
+ else:
86
+ x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
87
+ return x
88
+
89
+
90
+ class ResnetBlock(nn.Module):
91
+ def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
92
+ dropout, temb_channels=512):
93
+ super().__init__()
94
+ self.in_channels = in_channels
95
+ out_channels = in_channels if out_channels is None else out_channels
96
+ self.out_channels = out_channels
97
+ self.use_conv_shortcut = conv_shortcut
98
+
99
+ self.norm1 = Normalize(in_channels)
100
+ self.conv1 = torch.nn.Conv2d(in_channels,
101
+ out_channels,
102
+ kernel_size=3,
103
+ stride=1,
104
+ padding=1)
105
+ if temb_channels > 0:
106
+ self.temb_proj = torch.nn.Linear(temb_channels,
107
+ out_channels)
108
+ self.norm2 = Normalize(out_channels)
109
+ self.dropout = torch.nn.Dropout(dropout)
110
+ self.conv2 = torch.nn.Conv2d(out_channels,
111
+ out_channels,
112
+ kernel_size=3,
113
+ stride=1,
114
+ padding=1)
115
+ if self.in_channels != self.out_channels:
116
+ if self.use_conv_shortcut:
117
+ self.conv_shortcut = torch.nn.Conv2d(in_channels,
118
+ out_channels,
119
+ kernel_size=3,
120
+ stride=1,
121
+ padding=1)
122
+ else:
123
+ self.nin_shortcut = torch.nn.Conv2d(in_channels,
124
+ out_channels,
125
+ kernel_size=1,
126
+ stride=1,
127
+ padding=0)
128
+
129
+ def forward(self, x, temb):
130
+ h = x
131
+ h = self.norm1(h)
132
+ h = nonlinearity(h)
133
+ h = self.conv1(h)
134
+
135
+ if temb is not None:
136
+ h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
137
+
138
+ h = self.norm2(h)
139
+ h = nonlinearity(h)
140
+ h = self.dropout(h)
141
+ h = self.conv2(h)
142
+
143
+ if self.in_channels != self.out_channels:
144
+ if self.use_conv_shortcut:
145
+ x = self.conv_shortcut(x)
146
+ else:
147
+ x = self.nin_shortcut(x)
148
+
149
+ return x+h
150
+
151
+
152
+ class AttnBlock(nn.Module):
153
+ def __init__(self, in_channels):
154
+ super().__init__()
155
+ self.in_channels = in_channels
156
+
157
+ self.norm = Normalize(in_channels)
158
+ self.q = torch.nn.Conv2d(in_channels,
159
+ in_channels,
160
+ kernel_size=1,
161
+ stride=1,
162
+ padding=0)
163
+ self.k = torch.nn.Conv2d(in_channels,
164
+ in_channels,
165
+ kernel_size=1,
166
+ stride=1,
167
+ padding=0)
168
+ self.v = torch.nn.Conv2d(in_channels,
169
+ in_channels,
170
+ kernel_size=1,
171
+ stride=1,
172
+ padding=0)
173
+ self.proj_out = torch.nn.Conv2d(in_channels,
174
+ in_channels,
175
+ kernel_size=1,
176
+ stride=1,
177
+ padding=0)
178
+
179
+ def forward(self, x):
180
+ h_ = x
181
+ h_ = self.norm(h_)
182
+ q = self.q(h_)
183
+ k = self.k(h_)
184
+ v = self.v(h_)
185
+
186
+ # compute attention
187
+ b,c,h,w = q.shape
188
+ q = q.reshape(b,c,h*w)
189
+ q = q.permute(0,2,1) # b,hw,c
190
+ k = k.reshape(b,c,h*w) # b,c,hw
191
+ w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
192
+ w_ = w_ * (int(c)**(-0.5))
193
+ w_ = torch.nn.functional.softmax(w_, dim=2)
194
+
195
+ # attend to values
196
+ v = v.reshape(b,c,h*w)
197
+ w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q)
198
+ h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
199
+ h_ = h_.reshape(b,c,h,w)
200
+
201
+ h_ = self.proj_out(h_)
202
+
203
+ return x+h_
204
+
205
+ class MemoryEfficientAttnBlock(nn.Module):
206
+ """
207
+ Uses xformers efficient implementation,
208
+ see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
209
+ Note: this is a single-head self-attention operation
210
+ """
211
+ #
212
+ def __init__(self, in_channels):
213
+ super().__init__()
214
+ self.in_channels = in_channels
215
+
216
+ self.norm = Normalize(in_channels)
217
+ self.q = torch.nn.Conv2d(in_channels,
218
+ in_channels,
219
+ kernel_size=1,
220
+ stride=1,
221
+ padding=0)
222
+ self.k = torch.nn.Conv2d(in_channels,
223
+ in_channels,
224
+ kernel_size=1,
225
+ stride=1,
226
+ padding=0)
227
+ self.v = torch.nn.Conv2d(in_channels,
228
+ in_channels,
229
+ kernel_size=1,
230
+ stride=1,
231
+ padding=0)
232
+ self.proj_out = torch.nn.Conv2d(in_channels,
233
+ in_channels,
234
+ kernel_size=1,
235
+ stride=1,
236
+ padding=0)
237
+ self.attention_op: Optional[Any] = None
238
+
239
+ def forward(self, x):
240
+ h_ = x
241
+ h_ = self.norm(h_)
242
+ q = self.q(h_)
243
+ k = self.k(h_)
244
+ v = self.v(h_)
245
+
246
+ # compute attention
247
+ B, C, H, W = q.shape
248
+ q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))
249
+
250
+ q, k, v = map(
251
+ lambda t: t.unsqueeze(3)
252
+ .reshape(B, t.shape[1], 1, C)
253
+ .permute(0, 2, 1, 3)
254
+ .reshape(B * 1, t.shape[1], C)
255
+ .contiguous(),
256
+ (q, k, v),
257
+ )
258
+ out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
259
+
260
+ out = (
261
+ out.unsqueeze(0)
262
+ .reshape(B, 1, out.shape[1], C)
263
+ .permute(0, 2, 1, 3)
264
+ .reshape(B, out.shape[1], C)
265
+ )
266
+ out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
267
+ out = self.proj_out(out)
268
+ return x+out
269
+
270
+
271
+ class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
272
+ def forward(self, x, context=None, mask=None):
273
+ b, c, h, w = x.shape
274
+ x = rearrange(x, 'b c h w -> b (h w) c')
275
+ out = super().forward(x, context=context, mask=mask)
276
+ out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c)
277
+ return x + out
278
+
279
+
280
+ def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
281
+ assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear", "none"], f'attn_type {attn_type} unknown'
282
+ if XFORMERS_IS_AVAILBLE and attn_type == "vanilla":
283
+ attn_type = "vanilla-xformers"
284
+ print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
285
+ if attn_type == "vanilla":
286
+ assert attn_kwargs is None
287
+ return AttnBlock(in_channels)
288
+ elif attn_type == "vanilla-xformers":
289
+ print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
290
+ return MemoryEfficientAttnBlock(in_channels)
291
+ elif type == "memory-efficient-cross-attn":
292
+ attn_kwargs["query_dim"] = in_channels
293
+ return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
294
+ elif attn_type == "none":
295
+ return nn.Identity(in_channels)
296
+ else:
297
+ raise NotImplementedError()
298
+
299
+
300
+ class Model(nn.Module):
301
+ def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
302
+ attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
303
+ resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
304
+ super().__init__()
305
+ if use_linear_attn: attn_type = "linear"
306
+ self.ch = ch
307
+ self.temb_ch = self.ch*4
308
+ self.num_resolutions = len(ch_mult)
309
+ self.num_res_blocks = num_res_blocks
310
+ self.resolution = resolution
311
+ self.in_channels = in_channels
312
+
313
+ self.use_timestep = use_timestep
314
+ if self.use_timestep:
315
+ # timestep embedding
316
+ self.temb = nn.Module()
317
+ self.temb.dense = nn.ModuleList([
318
+ torch.nn.Linear(self.ch,
319
+ self.temb_ch),
320
+ torch.nn.Linear(self.temb_ch,
321
+ self.temb_ch),
322
+ ])
323
+
324
+ # downsampling
325
+ self.conv_in = torch.nn.Conv2d(in_channels,
326
+ self.ch,
327
+ kernel_size=3,
328
+ stride=1,
329
+ padding=1)
330
+
331
+ curr_res = resolution
332
+ in_ch_mult = (1,)+tuple(ch_mult)
333
+ self.down = nn.ModuleList()
334
+ for i_level in range(self.num_resolutions):
335
+ block = nn.ModuleList()
336
+ attn = nn.ModuleList()
337
+ block_in = ch*in_ch_mult[i_level]
338
+ block_out = ch*ch_mult[i_level]
339
+ for i_block in range(self.num_res_blocks):
340
+ block.append(ResnetBlock(in_channels=block_in,
341
+ out_channels=block_out,
342
+ temb_channels=self.temb_ch,
343
+ dropout=dropout))
344
+ block_in = block_out
345
+ if curr_res in attn_resolutions:
346
+ attn.append(make_attn(block_in, attn_type=attn_type))
347
+ down = nn.Module()
348
+ down.block = block
349
+ down.attn = attn
350
+ if i_level != self.num_resolutions-1:
351
+ down.downsample = Downsample(block_in, resamp_with_conv)
352
+ curr_res = curr_res // 2
353
+ self.down.append(down)
354
+
355
+ # middle
356
+ self.mid = nn.Module()
357
+ self.mid.block_1 = ResnetBlock(in_channels=block_in,
358
+ out_channels=block_in,
359
+ temb_channels=self.temb_ch,
360
+ dropout=dropout)
361
+ self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
362
+ self.mid.block_2 = ResnetBlock(in_channels=block_in,
363
+ out_channels=block_in,
364
+ temb_channels=self.temb_ch,
365
+ dropout=dropout)
366
+
367
+ # upsampling
368
+ self.up = nn.ModuleList()
369
+ for i_level in reversed(range(self.num_resolutions)):
370
+ block = nn.ModuleList()
371
+ attn = nn.ModuleList()
372
+ block_out = ch*ch_mult[i_level]
373
+ skip_in = ch*ch_mult[i_level]
374
+ for i_block in range(self.num_res_blocks+1):
375
+ if i_block == self.num_res_blocks:
376
+ skip_in = ch*in_ch_mult[i_level]
377
+ block.append(ResnetBlock(in_channels=block_in+skip_in,
378
+ out_channels=block_out,
379
+ temb_channels=self.temb_ch,
380
+ dropout=dropout))
381
+ block_in = block_out
382
+ if curr_res in attn_resolutions:
383
+ attn.append(make_attn(block_in, attn_type=attn_type))
384
+ up = nn.Module()
385
+ up.block = block
386
+ up.attn = attn
387
+ if i_level != 0:
388
+ up.upsample = Upsample(block_in, resamp_with_conv)
389
+ curr_res = curr_res * 2
390
+ self.up.insert(0, up) # prepend to get consistent order
391
+
392
+ # end
393
+ self.norm_out = Normalize(block_in)
394
+ self.conv_out = torch.nn.Conv2d(block_in,
395
+ out_ch,
396
+ kernel_size=3,
397
+ stride=1,
398
+ padding=1)
399
+
400
+ def forward(self, x, t=None, context=None):
401
+ #assert x.shape[2] == x.shape[3] == self.resolution
402
+ if context is not None:
403
+ # assume aligned context, cat along channel axis
404
+ x = torch.cat((x, context), dim=1)
405
+ if self.use_timestep:
406
+ # timestep embedding
407
+ assert t is not None
408
+ temb = get_timestep_embedding(t, self.ch)
409
+ temb = self.temb.dense[0](temb)
410
+ temb = nonlinearity(temb)
411
+ temb = self.temb.dense[1](temb)
412
+ else:
413
+ temb = None
414
+
415
+ # downsampling
416
+ hs = [self.conv_in(x)]
417
+ for i_level in range(self.num_resolutions):
418
+ for i_block in range(self.num_res_blocks):
419
+ h = self.down[i_level].block[i_block](hs[-1], temb)
420
+ if len(self.down[i_level].attn) > 0:
421
+ h = self.down[i_level].attn[i_block](h)
422
+ hs.append(h)
423
+ if i_level != self.num_resolutions-1:
424
+ hs.append(self.down[i_level].downsample(hs[-1]))
425
+
426
+ # middle
427
+ h = hs[-1]
428
+ h = self.mid.block_1(h, temb)
429
+ h = self.mid.attn_1(h)
430
+ h = self.mid.block_2(h, temb)
431
+
432
+ # upsampling
433
+ for i_level in reversed(range(self.num_resolutions)):
434
+ for i_block in range(self.num_res_blocks+1):
435
+ h = self.up[i_level].block[i_block](
436
+ torch.cat([h, hs.pop()], dim=1), temb)
437
+ if len(self.up[i_level].attn) > 0:
438
+ h = self.up[i_level].attn[i_block](h)
439
+ if i_level != 0:
440
+ h = self.up[i_level].upsample(h)
441
+
442
+ # end
443
+ h = self.norm_out(h)
444
+ h = nonlinearity(h)
445
+ h = self.conv_out(h)
446
+ return h
447
+
448
+ def get_last_layer(self):
449
+ return self.conv_out.weight
450
+
451
+
452
+ class Encoder(nn.Module):
453
+ def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
454
+ attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
455
+ resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
456
+ **ignore_kwargs):
457
+ super().__init__()
458
+ if use_linear_attn: attn_type = "linear"
459
+ self.ch = ch
460
+ self.temb_ch = 0
461
+ self.num_resolutions = len(ch_mult)
462
+ self.num_res_blocks = num_res_blocks
463
+ self.resolution = resolution
464
+ self.in_channels = in_channels
465
+
466
+ # downsampling
467
+ self.conv_in = torch.nn.Conv2d(in_channels,
468
+ self.ch,
469
+ kernel_size=3,
470
+ stride=1,
471
+ padding=1)
472
+
473
+ curr_res = resolution
474
+ in_ch_mult = (1,)+tuple(ch_mult)
475
+ self.in_ch_mult = in_ch_mult
476
+ self.down = nn.ModuleList()
477
+ for i_level in range(self.num_resolutions):
478
+ block = nn.ModuleList()
479
+ attn = nn.ModuleList()
480
+ block_in = ch*in_ch_mult[i_level]
481
+ block_out = ch*ch_mult[i_level]
482
+ for i_block in range(self.num_res_blocks):
483
+ block.append(ResnetBlock(in_channels=block_in,
484
+ out_channels=block_out,
485
+ temb_channels=self.temb_ch,
486
+ dropout=dropout))
487
+ block_in = block_out
488
+ if curr_res in attn_resolutions:
489
+ attn.append(make_attn(block_in, attn_type=attn_type))
490
+ down = nn.Module()
491
+ down.block = block
492
+ down.attn = attn
493
+ if i_level != self.num_resolutions-1:
494
+ down.downsample = Downsample(block_in, resamp_with_conv)
495
+ curr_res = curr_res // 2
496
+ self.down.append(down)
497
+
498
+ # middle
499
+ self.mid = nn.Module()
500
+ self.mid.block_1 = ResnetBlock(in_channels=block_in,
501
+ out_channels=block_in,
502
+ temb_channels=self.temb_ch,
503
+ dropout=dropout)
504
+ self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
505
+ self.mid.block_2 = ResnetBlock(in_channels=block_in,
506
+ out_channels=block_in,
507
+ temb_channels=self.temb_ch,
508
+ dropout=dropout)
509
+
510
+ # end
511
+ self.norm_out = Normalize(block_in)
512
+ self.conv_out = torch.nn.Conv2d(block_in,
513
+ 2*z_channels if double_z else z_channels,
514
+ kernel_size=3,
515
+ stride=1,
516
+ padding=1)
517
+
518
+ def forward(self, x):
519
+ # timestep embedding
520
+ temb = None
521
+
522
+ # downsampling
523
+ hs = [self.conv_in(x)]
524
+ for i_level in range(self.num_resolutions):
525
+ for i_block in range(self.num_res_blocks):
526
+ h = self.down[i_level].block[i_block](hs[-1], temb)
527
+ if len(self.down[i_level].attn) > 0:
528
+ h = self.down[i_level].attn[i_block](h)
529
+ hs.append(h)
530
+ if i_level != self.num_resolutions-1:
531
+ hs.append(self.down[i_level].downsample(hs[-1]))
532
+
533
+ # middle
534
+ h = hs[-1]
535
+ h = self.mid.block_1(h, temb)
536
+ h = self.mid.attn_1(h)
537
+ h = self.mid.block_2(h, temb)
538
+
539
+ # end
540
+ h = self.norm_out(h)
541
+ h = nonlinearity(h)
542
+ h = self.conv_out(h)
543
+ return h
544
+
545
+
546
+ class Decoder(nn.Module):
547
+ def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
548
+ attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
549
+ resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
550
+ attn_type="vanilla", **ignorekwargs):
551
+ super().__init__()
552
+ if use_linear_attn: attn_type = "linear"
553
+ self.ch = ch
554
+ self.temb_ch = 0
555
+ self.num_resolutions = len(ch_mult)
556
+ self.num_res_blocks = num_res_blocks
557
+ self.resolution = resolution
558
+ self.in_channels = in_channels
559
+ self.give_pre_end = give_pre_end
560
+ self.tanh_out = tanh_out
561
+
562
+ # compute in_ch_mult, block_in and curr_res at lowest res
563
+ in_ch_mult = (1,)+tuple(ch_mult)
564
+ block_in = ch*ch_mult[self.num_resolutions-1]
565
+ curr_res = resolution // 2**(self.num_resolutions-1)
566
+ self.z_shape = (1,z_channels,curr_res,curr_res)
567
+ print("Working with z of shape {} = {} dimensions.".format(
568
+ self.z_shape, np.prod(self.z_shape)))
569
+
570
+ # z to block_in
571
+ self.conv_in = torch.nn.Conv2d(z_channels,
572
+ block_in,
573
+ kernel_size=3,
574
+ stride=1,
575
+ padding=1)
576
+
577
+ # middle
578
+ self.mid = nn.Module()
579
+ self.mid.block_1 = ResnetBlock(in_channels=block_in,
580
+ out_channels=block_in,
581
+ temb_channels=self.temb_ch,
582
+ dropout=dropout)
583
+ self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
584
+ self.mid.block_2 = ResnetBlock(in_channels=block_in,
585
+ out_channels=block_in,
586
+ temb_channels=self.temb_ch,
587
+ dropout=dropout)
588
+
589
+ # upsampling
590
+ self.up = nn.ModuleList()
591
+ for i_level in reversed(range(self.num_resolutions)):
592
+ block = nn.ModuleList()
593
+ attn = nn.ModuleList()
594
+ block_out = ch*ch_mult[i_level]
595
+ for i_block in range(self.num_res_blocks+1):
596
+ block.append(ResnetBlock(in_channels=block_in,
597
+ out_channels=block_out,
598
+ temb_channels=self.temb_ch,
599
+ dropout=dropout))
600
+ block_in = block_out
601
+ if curr_res in attn_resolutions:
602
+ attn.append(make_attn(block_in, attn_type=attn_type))
603
+ up = nn.Module()
604
+ up.block = block
605
+ up.attn = attn
606
+ if i_level != 0:
607
+ up.upsample = Upsample(block_in, resamp_with_conv)
608
+ curr_res = curr_res * 2
609
+ self.up.insert(0, up) # prepend to get consistent order
610
+
611
+ # end
612
+ self.norm_out = Normalize(block_in)
613
+ self.conv_out = torch.nn.Conv2d(block_in,
614
+ out_ch,
615
+ kernel_size=3,
616
+ stride=1,
617
+ padding=1)
618
+
619
+ def forward(self, z):
620
+ #assert z.shape[1:] == self.z_shape[1:]
621
+ self.last_z_shape = z.shape
622
+
623
+ # timestep embedding
624
+ temb = None
625
+
626
+ # z to block_in
627
+ h = self.conv_in(z)
628
+
629
+ # middle
630
+ h = self.mid.block_1(h, temb)
631
+ h = self.mid.attn_1(h)
632
+ h = self.mid.block_2(h, temb)
633
+
634
+ # upsampling
635
+ for i_level in reversed(range(self.num_resolutions)):
636
+ for i_block in range(self.num_res_blocks+1):
637
+ h = self.up[i_level].block[i_block](h, temb)
638
+ if len(self.up[i_level].attn) > 0:
639
+ h = self.up[i_level].attn[i_block](h)
640
+ if i_level != 0:
641
+ h = self.up[i_level].upsample(h)
642
+
643
+ # end
644
+ if self.give_pre_end:
645
+ return h
646
+
647
+ h = self.norm_out(h)
648
+ h = nonlinearity(h)
649
+ h = self.conv_out(h)
650
+ if self.tanh_out:
651
+ h = torch.tanh(h)
652
+ return h
653
+
654
+
655
+ class SimpleDecoder(nn.Module):
656
+ def __init__(self, in_channels, out_channels, *args, **kwargs):
657
+ super().__init__()
658
+ self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
659
+ ResnetBlock(in_channels=in_channels,
660
+ out_channels=2 * in_channels,
661
+ temb_channels=0, dropout=0.0),
662
+ ResnetBlock(in_channels=2 * in_channels,
663
+ out_channels=4 * in_channels,
664
+ temb_channels=0, dropout=0.0),
665
+ ResnetBlock(in_channels=4 * in_channels,
666
+ out_channels=2 * in_channels,
667
+ temb_channels=0, dropout=0.0),
668
+ nn.Conv2d(2*in_channels, in_channels, 1),
669
+ Upsample(in_channels, with_conv=True)])
670
+ # end
671
+ self.norm_out = Normalize(in_channels)
672
+ self.conv_out = torch.nn.Conv2d(in_channels,
673
+ out_channels,
674
+ kernel_size=3,
675
+ stride=1,
676
+ padding=1)
677
+
678
+ def forward(self, x):
679
+ for i, layer in enumerate(self.model):
680
+ if i in [1,2,3]:
681
+ x = layer(x, None)
682
+ else:
683
+ x = layer(x)
684
+
685
+ h = self.norm_out(x)
686
+ h = nonlinearity(h)
687
+ x = self.conv_out(h)
688
+ return x
689
+
690
+
691
+ class UpsampleDecoder(nn.Module):
692
+ def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
693
+ ch_mult=(2,2), dropout=0.0):
694
+ super().__init__()
695
+ # upsampling
696
+ self.temb_ch = 0
697
+ self.num_resolutions = len(ch_mult)
698
+ self.num_res_blocks = num_res_blocks
699
+ block_in = in_channels
700
+ curr_res = resolution // 2 ** (self.num_resolutions - 1)
701
+ self.res_blocks = nn.ModuleList()
702
+ self.upsample_blocks = nn.ModuleList()
703
+ for i_level in range(self.num_resolutions):
704
+ res_block = []
705
+ block_out = ch * ch_mult[i_level]
706
+ for i_block in range(self.num_res_blocks + 1):
707
+ res_block.append(ResnetBlock(in_channels=block_in,
708
+ out_channels=block_out,
709
+ temb_channels=self.temb_ch,
710
+ dropout=dropout))
711
+ block_in = block_out
712
+ self.res_blocks.append(nn.ModuleList(res_block))
713
+ if i_level != self.num_resolutions - 1:
714
+ self.upsample_blocks.append(Upsample(block_in, True))
715
+ curr_res = curr_res * 2
716
+
717
+ # end
718
+ self.norm_out = Normalize(block_in)
719
+ self.conv_out = torch.nn.Conv2d(block_in,
720
+ out_channels,
721
+ kernel_size=3,
722
+ stride=1,
723
+ padding=1)
724
+
725
+ def forward(self, x):
726
+ # upsampling
727
+ h = x
728
+ for k, i_level in enumerate(range(self.num_resolutions)):
729
+ for i_block in range(self.num_res_blocks + 1):
730
+ h = self.res_blocks[i_level][i_block](h, None)
731
+ if i_level != self.num_resolutions - 1:
732
+ h = self.upsample_blocks[k](h)
733
+ h = self.norm_out(h)
734
+ h = nonlinearity(h)
735
+ h = self.conv_out(h)
736
+ return h
737
+
738
+
739
+ class LatentRescaler(nn.Module):
740
+ def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
741
+ super().__init__()
742
+ # residual block, interpolate, residual block
743
+ self.factor = factor
744
+ self.conv_in = nn.Conv2d(in_channels,
745
+ mid_channels,
746
+ kernel_size=3,
747
+ stride=1,
748
+ padding=1)
749
+ self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
750
+ out_channels=mid_channels,
751
+ temb_channels=0,
752
+ dropout=0.0) for _ in range(depth)])
753
+ self.attn = AttnBlock(mid_channels)
754
+ self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
755
+ out_channels=mid_channels,
756
+ temb_channels=0,
757
+ dropout=0.0) for _ in range(depth)])
758
+
759
+ self.conv_out = nn.Conv2d(mid_channels,
760
+ out_channels,
761
+ kernel_size=1,
762
+ )
763
+
764
+ def forward(self, x):
765
+ x = self.conv_in(x)
766
+ for block in self.res_block1:
767
+ x = block(x, None)
768
+ x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
769
+ x = self.attn(x)
770
+ for block in self.res_block2:
771
+ x = block(x, None)
772
+ x = self.conv_out(x)
773
+ return x
774
+
775
+
776
+ class MergedRescaleEncoder(nn.Module):
777
+ def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
778
+ attn_resolutions, dropout=0.0, resamp_with_conv=True,
779
+ ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
780
+ super().__init__()
781
+ intermediate_chn = ch * ch_mult[-1]
782
+ self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
783
+ z_channels=intermediate_chn, double_z=False, resolution=resolution,
784
+ attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
785
+ out_ch=None)
786
+ self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
787
+ mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
788
+
789
+ def forward(self, x):
790
+ x = self.encoder(x)
791
+ x = self.rescaler(x)
792
+ return x
793
+
794
+
795
+ class MergedRescaleDecoder(nn.Module):
796
+ def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
797
+ dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
798
+ super().__init__()
799
+ tmp_chn = z_channels*ch_mult[-1]
800
+ self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
801
+ resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
802
+ ch_mult=ch_mult, resolution=resolution, ch=ch)
803
+ self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
804
+ out_channels=tmp_chn, depth=rescale_module_depth)
805
+
806
+ def forward(self, x):
807
+ x = self.rescaler(x)
808
+ x = self.decoder(x)
809
+ return x
810
+
811
+
812
+ class Upsampler(nn.Module):
813
+ def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
814
+ super().__init__()
815
+ assert out_size >= in_size
816
+ num_blocks = int(np.log2(out_size//in_size))+1
817
+ factor_up = 1.+ (out_size % in_size)
818
+ print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
819
+ self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
820
+ out_channels=in_channels)
821
+ self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
822
+ attn_resolutions=[], in_channels=None, ch=in_channels,
823
+ ch_mult=[ch_mult for _ in range(num_blocks)])
824
+
825
+ def forward(self, x):
826
+ x = self.rescaler(x)
827
+ x = self.decoder(x)
828
+ return x
829
+
830
+
831
+ class Resize(nn.Module):
832
+ def __init__(self, in_channels=None, learned=False, mode="bilinear"):
833
+ super().__init__()
834
+ self.with_conv = learned
835
+ self.mode = mode
836
+ if self.with_conv:
837
+ print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
838
+ raise NotImplementedError()
839
+ assert in_channels is not None
840
+ # no asymmetric padding in torch conv, must do it ourselves
841
+ self.conv = torch.nn.Conv2d(in_channels,
842
+ in_channels,
843
+ kernel_size=4,
844
+ stride=2,
845
+ padding=1)
846
+
847
+ def forward(self, x, scale_factor=1.0):
848
+ if scale_factor==1.0:
849
+ return x
850
+ else:
851
+ x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
852
+ return x
ldm/modules/diffusionmodules/openaimodel.py ADDED
@@ -0,0 +1,790 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from abc import abstractmethod
2
+ import math
3
+
4
+ import numpy as np
5
+ import torch as th
6
+ import torch.nn as nn
7
+ import torch.nn.functional as F
8
+
9
+ from ldm.modules.diffusionmodules.util import (
10
+ checkpoint,
11
+ conv_nd,
12
+ linear,
13
+ avg_pool_nd,
14
+ zero_module,
15
+ normalization,
16
+ timestep_embedding,
17
+ )
18
+ from ldm.modules.attention import SpatialTransformer
19
+ from ldm.util import exists
20
+
21
+
22
+ # dummy replace
23
+ def convert_module_to_f16(x):
24
+ pass
25
+
26
+ def convert_module_to_f32(x):
27
+ pass
28
+
29
+
30
+ ## go
31
+ class AttentionPool2d(nn.Module):
32
+ """
33
+ Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
34
+ """
35
+
36
+ def __init__(
37
+ self,
38
+ spacial_dim: int,
39
+ embed_dim: int,
40
+ num_heads_channels: int,
41
+ output_dim: int = None,
42
+ ):
43
+ super().__init__()
44
+ self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
45
+ self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
46
+ self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
47
+ self.num_heads = embed_dim // num_heads_channels
48
+ self.attention = QKVAttention(self.num_heads)
49
+
50
+ def forward(self, x):
51
+ b, c, *_spatial = x.shape
52
+ x = x.reshape(b, c, -1) # NC(HW)
53
+ x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1)
54
+ x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1)
55
+ x = self.qkv_proj(x)
56
+ x = self.attention(x)
57
+ x = self.c_proj(x)
58
+ return x[:, :, 0]
59
+
60
+
61
+ class TimestepBlock(nn.Module):
62
+ """
63
+ Any module where forward() takes timestep embeddings as a second argument.
64
+ """
65
+
66
+ @abstractmethod
67
+ def forward(self, x, emb):
68
+ """
69
+ Apply the module to `x` given `emb` timestep embeddings.
70
+ """
71
+
72
+
73
+ class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
74
+ """
75
+ A sequential module that passes timestep embeddings to the children that
76
+ support it as an extra input.
77
+ """
78
+
79
+ def forward(self, x, emb, context=None, hint=None):
80
+ for layer in self:
81
+ if isinstance(layer, TimestepBlock):
82
+ x = layer(x, emb)
83
+ elif isinstance(layer, SpatialTransformer):
84
+ x = layer(x, context, hint)
85
+ else:
86
+ x = layer(x)
87
+ return x
88
+
89
+
90
+ class Upsample(nn.Module):
91
+ """
92
+ An upsampling layer with an optional convolution.
93
+ :param channels: channels in the inputs and outputs.
94
+ :param use_conv: a bool determining if a convolution is applied.
95
+ :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
96
+ upsampling occurs in the inner-two dimensions.
97
+ """
98
+
99
+ def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
100
+ super().__init__()
101
+ self.channels = channels
102
+ self.out_channels = out_channels or channels
103
+ self.use_conv = use_conv
104
+ self.dims = dims
105
+ if use_conv:
106
+ self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
107
+
108
+ def forward(self, x):
109
+ assert x.shape[1] == self.channels
110
+ if self.dims == 3:
111
+ x = F.interpolate(
112
+ x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
113
+ )
114
+ else:
115
+ x = F.interpolate(x, scale_factor=2, mode="nearest")
116
+ if self.use_conv:
117
+ x = self.conv(x)
118
+ return x
119
+
120
+ class TransposedUpsample(nn.Module):
121
+ 'Learned 2x upsampling without padding'
122
+ def __init__(self, channels, out_channels=None, ks=5):
123
+ super().__init__()
124
+ self.channels = channels
125
+ self.out_channels = out_channels or channels
126
+
127
+ self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
128
+
129
+ def forward(self,x):
130
+ return self.up(x)
131
+
132
+
133
+ class Downsample(nn.Module):
134
+ """
135
+ A downsampling layer with an optional convolution.
136
+ :param channels: channels in the inputs and outputs.
137
+ :param use_conv: a bool determining if a convolution is applied.
138
+ :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
139
+ downsampling occurs in the inner-two dimensions.
140
+ """
141
+
142
+ def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
143
+ super().__init__()
144
+ self.channels = channels
145
+ self.out_channels = out_channels or channels
146
+ self.use_conv = use_conv
147
+ self.dims = dims
148
+ stride = 2 if dims != 3 else (1, 2, 2)
149
+ if use_conv:
150
+ self.op = conv_nd(
151
+ dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
152
+ )
153
+ else:
154
+ assert self.channels == self.out_channels
155
+ self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
156
+
157
+ def forward(self, x):
158
+ assert x.shape[1] == self.channels
159
+ return self.op(x)
160
+
161
+
162
+ class ResBlock(TimestepBlock):
163
+ """
164
+ A residual block that can optionally change the number of channels.
165
+ :param channels: the number of input channels.
166
+ :param emb_channels: the number of timestep embedding channels.
167
+ :param dropout: the rate of dropout.
168
+ :param out_channels: if specified, the number of out channels.
169
+ :param use_conv: if True and out_channels is specified, use a spatial
170
+ convolution instead of a smaller 1x1 convolution to change the
171
+ channels in the skip connection.
172
+ :param dims: determines if the signal is 1D, 2D, or 3D.
173
+ :param use_checkpoint: if True, use gradient checkpointing on this module.
174
+ :param up: if True, use this block for upsampling.
175
+ :param down: if True, use this block for downsampling.
176
+ """
177
+
178
+ def __init__(
179
+ self,
180
+ channels,
181
+ emb_channels,
182
+ dropout,
183
+ out_channels=None,
184
+ use_conv=False,
185
+ use_scale_shift_norm=False,
186
+ dims=2,
187
+ use_checkpoint=False,
188
+ up=False,
189
+ down=False,
190
+ ):
191
+ super().__init__()
192
+ self.channels = channels
193
+ self.emb_channels = emb_channels
194
+ self.dropout = dropout
195
+ self.out_channels = out_channels or channels
196
+ self.use_conv = use_conv
197
+ self.use_checkpoint = use_checkpoint
198
+ self.use_scale_shift_norm = use_scale_shift_norm
199
+
200
+ self.in_layers = nn.Sequential(
201
+ normalization(channels),
202
+ nn.SiLU(),
203
+ conv_nd(dims, channels, self.out_channels, 3, padding=1),
204
+ )
205
+
206
+ self.updown = up or down
207
+
208
+ if up:
209
+ self.h_upd = Upsample(channels, False, dims)
210
+ self.x_upd = Upsample(channels, False, dims)
211
+ elif down:
212
+ self.h_upd = Downsample(channels, False, dims)
213
+ self.x_upd = Downsample(channels, False, dims)
214
+ else:
215
+ self.h_upd = self.x_upd = nn.Identity()
216
+
217
+ self.emb_layers = nn.Sequential(
218
+ nn.SiLU(),
219
+ linear(
220
+ emb_channels,
221
+ 2 * self.out_channels if use_scale_shift_norm else self.out_channels,
222
+ ),
223
+ )
224
+ self.out_layers = nn.Sequential(
225
+ normalization(self.out_channels),
226
+ nn.SiLU(),
227
+ nn.Dropout(p=dropout),
228
+ zero_module(
229
+ conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
230
+ ),
231
+ )
232
+
233
+ if self.out_channels == channels:
234
+ self.skip_connection = nn.Identity()
235
+ elif use_conv:
236
+ self.skip_connection = conv_nd(
237
+ dims, channels, self.out_channels, 3, padding=1
238
+ )
239
+ else:
240
+ self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
241
+
242
+ def forward(self, x, emb):
243
+ """
244
+ Apply the block to a Tensor, conditioned on a timestep embedding.
245
+ :param x: an [N x C x ...] Tensor of features.
246
+ :param emb: an [N x emb_channels] Tensor of timestep embeddings.
247
+ :return: an [N x C x ...] Tensor of outputs.
248
+ """
249
+ return checkpoint(
250
+ self._forward, (x, emb), self.parameters(), self.use_checkpoint
251
+ )
252
+
253
+
254
+ def _forward(self, x, emb):
255
+ if self.updown:
256
+ in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
257
+ h = in_rest(x)
258
+ h = self.h_upd(h)
259
+ x = self.x_upd(x)
260
+ h = in_conv(h)
261
+ else:
262
+ h = self.in_layers(x)
263
+ emb_out = self.emb_layers(emb).type(h.dtype)
264
+ while len(emb_out.shape) < len(h.shape):
265
+ emb_out = emb_out[..., None]
266
+ if self.use_scale_shift_norm:
267
+ out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
268
+ scale, shift = th.chunk(emb_out, 2, dim=1)
269
+ h = out_norm(h) * (1 + scale) + shift
270
+ h = out_rest(h)
271
+ else:
272
+ h = h + emb_out
273
+ h = self.out_layers(h)
274
+ return self.skip_connection(x) + h
275
+
276
+
277
+ class AttentionBlock(nn.Module):
278
+ """
279
+ An attention block that allows spatial positions to attend to each other.
280
+ Originally ported from here, but adapted to the N-d case.
281
+ https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
282
+ """
283
+
284
+ def __init__(
285
+ self,
286
+ channels,
287
+ num_heads=1,
288
+ num_head_channels=-1,
289
+ use_checkpoint=False,
290
+ use_new_attention_order=False,
291
+ ):
292
+ super().__init__()
293
+ self.channels = channels
294
+ if num_head_channels == -1:
295
+ self.num_heads = num_heads
296
+ else:
297
+ assert (
298
+ channels % num_head_channels == 0
299
+ ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
300
+ self.num_heads = channels // num_head_channels
301
+ self.use_checkpoint = use_checkpoint
302
+ self.norm = normalization(channels)
303
+ self.qkv = conv_nd(1, channels, channels * 3, 1)
304
+ if use_new_attention_order:
305
+ # split qkv before split heads
306
+ self.attention = QKVAttention(self.num_heads)
307
+ else:
308
+ # split heads before split qkv
309
+ self.attention = QKVAttentionLegacy(self.num_heads)
310
+
311
+ self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
312
+
313
+ def forward(self, x):
314
+ return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
315
+ #return pt_checkpoint(self._forward, x) # pytorch
316
+
317
+ def _forward(self, x):
318
+ b, c, *spatial = x.shape
319
+ x = x.reshape(b, c, -1)
320
+ qkv = self.qkv(self.norm(x))
321
+ h = self.attention(qkv)
322
+ h = self.proj_out(h)
323
+ return (x + h).reshape(b, c, *spatial)
324
+
325
+
326
+ def count_flops_attn(model, _x, y):
327
+ """
328
+ A counter for the `thop` package to count the operations in an
329
+ attention operation.
330
+ Meant to be used like:
331
+ macs, params = thop.profile(
332
+ model,
333
+ inputs=(inputs, timestamps),
334
+ custom_ops={QKVAttention: QKVAttention.count_flops},
335
+ )
336
+ """
337
+ b, c, *spatial = y[0].shape
338
+ num_spatial = int(np.prod(spatial))
339
+ # We perform two matmuls with the same number of ops.
340
+ # The first computes the weight matrix, the second computes
341
+ # the combination of the value vectors.
342
+ matmul_ops = 2 * b * (num_spatial ** 2) * c
343
+ model.total_ops += th.DoubleTensor([matmul_ops])
344
+
345
+
346
+ class QKVAttentionLegacy(nn.Module):
347
+ """
348
+ A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
349
+ """
350
+
351
+ def __init__(self, n_heads):
352
+ super().__init__()
353
+ self.n_heads = n_heads
354
+
355
+ def forward(self, qkv):
356
+ """
357
+ Apply QKV attention.
358
+ :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
359
+ :return: an [N x (H * C) x T] tensor after attention.
360
+ """
361
+ bs, width, length = qkv.shape
362
+ assert width % (3 * self.n_heads) == 0
363
+ ch = width // (3 * self.n_heads)
364
+ q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
365
+ scale = 1 / math.sqrt(math.sqrt(ch))
366
+ weight = th.einsum(
367
+ "bct,bcs->bts", q * scale, k * scale
368
+ ) # More stable with f16 than dividing afterwards
369
+ weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
370
+ a = th.einsum("bts,bcs->bct", weight, v)
371
+ return a.reshape(bs, -1, length)
372
+
373
+ @staticmethod
374
+ def count_flops(model, _x, y):
375
+ return count_flops_attn(model, _x, y)
376
+
377
+
378
+ class QKVAttention(nn.Module):
379
+ """
380
+ A module which performs QKV attention and splits in a different order.
381
+ """
382
+
383
+ def __init__(self, n_heads):
384
+ super().__init__()
385
+ self.n_heads = n_heads
386
+
387
+ def forward(self, qkv):
388
+ """
389
+ Apply QKV attention.
390
+ :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
391
+ :return: an [N x (H * C) x T] tensor after attention.
392
+ """
393
+ bs, width, length = qkv.shape
394
+ assert width % (3 * self.n_heads) == 0
395
+ ch = width // (3 * self.n_heads)
396
+ q, k, v = qkv.chunk(3, dim=1)
397
+ scale = 1 / math.sqrt(math.sqrt(ch))
398
+ weight = th.einsum(
399
+ "bct,bcs->bts",
400
+ (q * scale).view(bs * self.n_heads, ch, length),
401
+ (k * scale).view(bs * self.n_heads, ch, length),
402
+ ) # More stable with f16 than dividing afterwards
403
+ weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
404
+ a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
405
+ return a.reshape(bs, -1, length)
406
+
407
+ @staticmethod
408
+ def count_flops(model, _x, y):
409
+ return count_flops_attn(model, _x, y)
410
+
411
+
412
+ class UNetModel(nn.Module):
413
+ """
414
+ The full UNet model with attention and timestep embedding.
415
+ :param in_channels: channels in the input Tensor.
416
+ :param model_channels: base channel count for the model.
417
+ :param out_channels: channels in the output Tensor.
418
+ :param num_res_blocks: number of residual blocks per downsample.
419
+ :param attention_resolutions: a collection of downsample rates at which
420
+ attention will take place. May be a set, list, or tuple.
421
+ For example, if this contains 4, then at 4x downsampling, attention
422
+ will be used.
423
+ :param dropout: the dropout probability.
424
+ :param channel_mult: channel multiplier for each level of the UNet.
425
+ :param conv_resample: if True, use learned convolutions for upsampling and
426
+ downsampling.
427
+ :param dims: determines if the signal is 1D, 2D, or 3D.
428
+ :param num_classes: if specified (as an int), then this model will be
429
+ class-conditional with `num_classes` classes.
430
+ :param use_checkpoint: use gradient checkpointing to reduce memory usage.
431
+ :param num_heads: the number of attention heads in each attention layer.
432
+ :param num_heads_channels: if specified, ignore num_heads and instead use
433
+ a fixed channel width per attention head.
434
+ :param num_heads_upsample: works with num_heads to set a different number
435
+ of heads for upsampling. Deprecated.
436
+ :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
437
+ :param resblock_updown: use residual blocks for up/downsampling.
438
+ :param use_new_attention_order: use a different attention pattern for potentially
439
+ increased efficiency.
440
+ """
441
+
442
+ def __init__(
443
+ self,
444
+ image_size,
445
+ in_channels,
446
+ model_channels,
447
+ out_channels,
448
+ num_res_blocks,
449
+ attention_resolutions,
450
+ dropout=0,
451
+ channel_mult=(1, 2, 4, 8),
452
+ conv_resample=True,
453
+ dims=2,
454
+ num_classes=None,
455
+ use_checkpoint=False,
456
+ use_fp16=False,
457
+ num_heads=-1,
458
+ num_head_channels=-1,
459
+ num_heads_upsample=-1,
460
+ use_scale_shift_norm=False,
461
+ resblock_updown=False,
462
+ use_new_attention_order=False,
463
+ use_spatial_transformer=False, # custom transformer support
464
+ transformer_depth=1, # custom transformer support
465
+ context_dim=None, # custom transformer support
466
+ n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
467
+ legacy=True,
468
+ disable_self_attentions=None,
469
+ num_attention_blocks=None,
470
+ disable_middle_self_attn=False,
471
+ use_linear_in_transformer=False,
472
+ no_control=False,
473
+ ):
474
+ super().__init__()
475
+ self.no_control = no_control
476
+ if use_spatial_transformer:
477
+ assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
478
+
479
+ if context_dim is not None:
480
+ assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
481
+ from omegaconf.listconfig import ListConfig
482
+ if type(context_dim) == ListConfig:
483
+ context_dim = list(context_dim)
484
+
485
+ if num_heads_upsample == -1:
486
+ num_heads_upsample = num_heads
487
+
488
+ if num_heads == -1:
489
+ assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
490
+
491
+ if num_head_channels == -1:
492
+ assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
493
+
494
+ self.image_size = image_size
495
+ self.in_channels = in_channels
496
+ self.model_channels = model_channels
497
+ self.out_channels = out_channels
498
+ if isinstance(num_res_blocks, int):
499
+ self.num_res_blocks = len(channel_mult) * [num_res_blocks]
500
+ else:
501
+ if len(num_res_blocks) != len(channel_mult):
502
+ raise ValueError("provide num_res_blocks either as an int (globally constant) or "
503
+ "as a list/tuple (per-level) with the same length as channel_mult")
504
+ self.num_res_blocks = num_res_blocks
505
+ if disable_self_attentions is not None:
506
+ # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
507
+ assert len(disable_self_attentions) == len(channel_mult)
508
+ if num_attention_blocks is not None:
509
+ assert len(num_attention_blocks) == len(self.num_res_blocks)
510
+ assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
511
+ print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
512
+ f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
513
+ f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
514
+ f"attention will still not be set.")
515
+
516
+ self.attention_resolutions = attention_resolutions
517
+ self.dropout = dropout
518
+ self.channel_mult = channel_mult
519
+ self.conv_resample = conv_resample
520
+ self.num_classes = num_classes
521
+ self.use_checkpoint = use_checkpoint
522
+ self.dtype = th.float16 if use_fp16 else th.float32
523
+ self.num_heads = num_heads
524
+ self.num_head_channels = num_head_channels
525
+ self.num_heads_upsample = num_heads_upsample
526
+ self.predict_codebook_ids = n_embed is not None
527
+ self.transformer_depth = transformer_depth
528
+ self.context_dim = context_dim
529
+ self.use_linear_in_transformer = use_linear_in_transformer
530
+
531
+ time_embed_dim = model_channels * 4
532
+ self.time_embed = nn.Sequential(
533
+ linear(model_channels, time_embed_dim),
534
+ nn.SiLU(),
535
+ linear(time_embed_dim, time_embed_dim),
536
+ )
537
+
538
+ if self.num_classes is not None:
539
+ if isinstance(self.num_classes, int):
540
+ self.label_emb = nn.Embedding(num_classes, time_embed_dim)
541
+ elif self.num_classes == "continuous":
542
+ print("setting up linear c_adm embedding layer")
543
+ self.label_emb = nn.Linear(1, time_embed_dim)
544
+ else:
545
+ raise ValueError()
546
+
547
+ self.input_blocks = nn.ModuleList(
548
+ [
549
+ TimestepEmbedSequential(
550
+ conv_nd(dims, in_channels, model_channels, 3, padding=1)
551
+ )
552
+ ]
553
+ )
554
+ self._feature_size = model_channels
555
+ input_block_chans = [model_channels]
556
+ ch = model_channels
557
+ ds = 1
558
+ for level, mult in enumerate(channel_mult):
559
+ for nr in range(self.num_res_blocks[level]):
560
+ layers = [
561
+ ResBlock(
562
+ ch,
563
+ time_embed_dim,
564
+ dropout,
565
+ out_channels=mult * model_channels,
566
+ dims=dims,
567
+ use_checkpoint=use_checkpoint,
568
+ use_scale_shift_norm=use_scale_shift_norm,
569
+ )
570
+ ]
571
+ ch = mult * model_channels
572
+ if ds in attention_resolutions:
573
+ if num_head_channels == -1:
574
+ dim_head = ch // num_heads
575
+ else:
576
+ num_heads = ch // num_head_channels
577
+ dim_head = num_head_channels
578
+ if legacy:
579
+ #num_heads = 1
580
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
581
+ if exists(disable_self_attentions):
582
+ disabled_sa = disable_self_attentions[level]
583
+ else:
584
+ disabled_sa = False
585
+
586
+ if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
587
+ layers.append(
588
+ AttentionBlock(
589
+ ch,
590
+ use_checkpoint=use_checkpoint,
591
+ num_heads=num_heads,
592
+ num_head_channels=dim_head,
593
+ use_new_attention_order=use_new_attention_order,
594
+ ) if not use_spatial_transformer else SpatialTransformer(
595
+ ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
596
+ disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
597
+ use_checkpoint=use_checkpoint
598
+ )
599
+ )
600
+ self.input_blocks.append(TimestepEmbedSequential(*layers))
601
+ self._feature_size += ch
602
+ input_block_chans.append(ch)
603
+ if level != len(channel_mult) - 1:
604
+ out_ch = ch
605
+ self.input_blocks.append(
606
+ TimestepEmbedSequential(
607
+ ResBlock(
608
+ ch,
609
+ time_embed_dim,
610
+ dropout,
611
+ out_channels=out_ch,
612
+ dims=dims,
613
+ use_checkpoint=use_checkpoint,
614
+ use_scale_shift_norm=use_scale_shift_norm,
615
+ down=True,
616
+ )
617
+ if resblock_updown
618
+ else Downsample(
619
+ ch, conv_resample, dims=dims, out_channels=out_ch
620
+ )
621
+ )
622
+ )
623
+ ch = out_ch
624
+ input_block_chans.append(ch)
625
+ ds *= 2
626
+ self._feature_size += ch
627
+
628
+ if num_head_channels == -1:
629
+ dim_head = ch // num_heads
630
+ else:
631
+ num_heads = ch // num_head_channels
632
+ dim_head = num_head_channels
633
+ if legacy:
634
+ #num_heads = 1
635
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
636
+ self.middle_block = TimestepEmbedSequential(
637
+ ResBlock(
638
+ ch,
639
+ time_embed_dim,
640
+ dropout,
641
+ dims=dims,
642
+ use_checkpoint=use_checkpoint,
643
+ use_scale_shift_norm=use_scale_shift_norm,
644
+ ),
645
+ AttentionBlock(
646
+ ch,
647
+ use_checkpoint=use_checkpoint,
648
+ num_heads=num_heads,
649
+ num_head_channels=dim_head,
650
+ use_new_attention_order=use_new_attention_order,
651
+ ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn
652
+ ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
653
+ disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
654
+ use_checkpoint=use_checkpoint
655
+ ),
656
+ ResBlock(
657
+ ch,
658
+ time_embed_dim,
659
+ dropout,
660
+ dims=dims,
661
+ use_checkpoint=use_checkpoint,
662
+ use_scale_shift_norm=use_scale_shift_norm,
663
+ ),
664
+ )
665
+ self._feature_size += ch
666
+
667
+ self.output_blocks = nn.ModuleList([])
668
+ for level, mult in list(enumerate(channel_mult))[::-1]:
669
+ for i in range(self.num_res_blocks[level] + 1):
670
+ ich = input_block_chans.pop()
671
+ layers = [
672
+ ResBlock(
673
+ ch + ich,
674
+ time_embed_dim,
675
+ dropout,
676
+ out_channels=model_channels * mult,
677
+ dims=dims,
678
+ use_checkpoint=use_checkpoint,
679
+ use_scale_shift_norm=use_scale_shift_norm,
680
+ )
681
+ ]
682
+ ch = model_channels * mult
683
+ if ds in attention_resolutions:
684
+ if num_head_channels == -1:
685
+ dim_head = ch // num_heads
686
+ else:
687
+ num_heads = ch // num_head_channels
688
+ dim_head = num_head_channels
689
+ if legacy:
690
+ #num_heads = 1
691
+ dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
692
+ if exists(disable_self_attentions):
693
+ disabled_sa = disable_self_attentions[level]
694
+ else:
695
+ disabled_sa = False
696
+
697
+ if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
698
+ layers.append(
699
+ AttentionBlock(
700
+ ch,
701
+ use_checkpoint=use_checkpoint,
702
+ num_heads=num_heads_upsample,
703
+ num_head_channels=dim_head,
704
+ use_new_attention_order=use_new_attention_order,
705
+ ) if not use_spatial_transformer else SpatialTransformer(
706
+ ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
707
+ disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
708
+ use_checkpoint=use_checkpoint
709
+ )
710
+ )
711
+ if level and i == self.num_res_blocks[level]:
712
+ out_ch = ch
713
+ layers.append(
714
+ ResBlock(
715
+ ch,
716
+ time_embed_dim,
717
+ dropout,
718
+ out_channels=out_ch,
719
+ dims=dims,
720
+ use_checkpoint=use_checkpoint,
721
+ use_scale_shift_norm=use_scale_shift_norm,
722
+ up=True,
723
+ )
724
+ if resblock_updown
725
+ else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
726
+ )
727
+ ds //= 2
728
+ self.output_blocks.append(TimestepEmbedSequential(*layers))
729
+ self._feature_size += ch
730
+
731
+ self.out = nn.Sequential(
732
+ normalization(ch),
733
+ nn.SiLU(),
734
+ zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
735
+ )
736
+ if self.predict_codebook_ids:
737
+ self.id_predictor = nn.Sequential(
738
+ normalization(ch),
739
+ conv_nd(dims, model_channels, n_embed, 1),
740
+ #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
741
+ )
742
+
743
+ def convert_to_fp16(self):
744
+ """
745
+ Convert the torso of the model to float16.
746
+ """
747
+ self.input_blocks.apply(convert_module_to_f16)
748
+ self.middle_block.apply(convert_module_to_f16)
749
+ self.output_blocks.apply(convert_module_to_f16)
750
+
751
+ def convert_to_fp32(self):
752
+ """
753
+ Convert the torso of the model to float32.
754
+ """
755
+ self.input_blocks.apply(convert_module_to_f32)
756
+ self.middle_block.apply(convert_module_to_f32)
757
+ self.output_blocks.apply(convert_module_to_f32)
758
+
759
+ def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
760
+ """
761
+ Apply the model to an input batch.
762
+ :param x: an [N x C x ...] Tensor of inputs.
763
+ :param timesteps: a 1-D batch of timesteps.
764
+ :param context: conditioning plugged in via crossattn
765
+ :param y: an [N] Tensor of labels, if class-conditional.
766
+ :return: an [N x C x ...] Tensor of outputs.
767
+ """
768
+ assert (y is not None) == (
769
+ self.num_classes is not None
770
+ ), "must specify y if and only if the model is class-conditional"
771
+ hs = []
772
+ t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
773
+ emb = self.time_embed(t_emb)
774
+ if self.num_classes is not None:
775
+ assert y.shape[0] == x.shape[0]
776
+ emb = emb + self.label_emb(y)
777
+
778
+ h = x.type(self.dtype)
779
+ for module in self.input_blocks:
780
+ h = module(h, emb, context)
781
+ hs.append(h)
782
+ h = self.middle_block(h, emb, context)
783
+ for module in self.output_blocks:
784
+ h = th.cat([h, hs.pop()], dim=1)
785
+ h = module(h, emb, context)
786
+ h = h.type(x.dtype)
787
+ if self.predict_codebook_ids:
788
+ return self.id_predictor(h)
789
+ else:
790
+ return self.out(h)
ldm/modules/diffusionmodules/upscaling.py ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ import numpy as np
4
+ from functools import partial
5
+
6
+ from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule
7
+ from ldm.util import default
8
+
9
+
10
+ class AbstractLowScaleModel(nn.Module):
11
+ # for concatenating a downsampled image to the latent representation
12
+ def __init__(self, noise_schedule_config=None):
13
+ super(AbstractLowScaleModel, self).__init__()
14
+ if noise_schedule_config is not None:
15
+ self.register_schedule(**noise_schedule_config)
16
+
17
+ def register_schedule(self, beta_schedule="linear", timesteps=1000,
18
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
19
+ betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
20
+ cosine_s=cosine_s)
21
+ alphas = 1. - betas
22
+ alphas_cumprod = np.cumprod(alphas, axis=0)
23
+ alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
24
+
25
+ timesteps, = betas.shape
26
+ self.num_timesteps = int(timesteps)
27
+ self.linear_start = linear_start
28
+ self.linear_end = linear_end
29
+ assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
30
+
31
+ to_torch = partial(torch.tensor, dtype=torch.float32)
32
+
33
+ self.register_buffer('betas', to_torch(betas))
34
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
35
+ self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
36
+
37
+ # calculations for diffusion q(x_t | x_{t-1}) and others
38
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
39
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
40
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
41
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
42
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
43
+
44
+ def q_sample(self, x_start, t, noise=None):
45
+ noise = default(noise, lambda: torch.randn_like(x_start))
46
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
47
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
48
+
49
+ def forward(self, x):
50
+ return x, None
51
+
52
+ def decode(self, x):
53
+ return x
54
+
55
+
56
+ class SimpleImageConcat(AbstractLowScaleModel):
57
+ # no noise level conditioning
58
+ def __init__(self):
59
+ super(SimpleImageConcat, self).__init__(noise_schedule_config=None)
60
+ self.max_noise_level = 0
61
+
62
+ def forward(self, x):
63
+ # fix to constant noise level
64
+ return x, torch.zeros(x.shape[0], device=x.device).long()
65
+
66
+
67
+ class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel):
68
+ def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False):
69
+ super().__init__(noise_schedule_config=noise_schedule_config)
70
+ self.max_noise_level = max_noise_level
71
+
72
+ def forward(self, x, noise_level=None):
73
+ if noise_level is None:
74
+ noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
75
+ else:
76
+ assert isinstance(noise_level, torch.Tensor)
77
+ z = self.q_sample(x, noise_level)
78
+ return z, noise_level
79
+
80
+
81
+
ldm/modules/diffusionmodules/util.py ADDED
@@ -0,0 +1,271 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # adopted from
2
+ # https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
3
+ # and
4
+ # https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
5
+ # and
6
+ # https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
7
+ #
8
+ # thanks!
9
+
10
+
11
+ import os
12
+ import math
13
+ import torch
14
+ import torch.nn as nn
15
+ import numpy as np
16
+ from einops import repeat
17
+
18
+ from ldm.util import instantiate_from_config
19
+
20
+
21
+ def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
22
+ print(f"beta scheduler name : {schedule}")
23
+ if schedule == "linear":
24
+ betas = (
25
+ torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
26
+ )
27
+
28
+ elif schedule == "cosine":
29
+ timesteps = (
30
+ torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
31
+ )
32
+ alphas = timesteps / (1 + cosine_s) * np.pi / 2
33
+ alphas = torch.cos(alphas).pow(2)
34
+ alphas = alphas / alphas[0]
35
+ betas = 1 - alphas[1:] / alphas[:-1]
36
+ betas = np.clip(betas, a_min=0, a_max=0.999)
37
+
38
+ elif schedule == "sqrt_linear":
39
+ betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
40
+ elif schedule == "sqrt":
41
+ betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
42
+ else:
43
+ raise ValueError(f"schedule '{schedule}' unknown.")
44
+ return betas.numpy()
45
+
46
+
47
+ def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
48
+ if ddim_discr_method == 'uniform':
49
+ c = num_ddpm_timesteps // num_ddim_timesteps
50
+ ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
51
+ elif ddim_discr_method == 'quad':
52
+ ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
53
+ else:
54
+ raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
55
+
56
+ # assert ddim_timesteps.shape[0] == num_ddim_timesteps
57
+ # add one to get the final alpha values right (the ones from first scale to data during sampling)
58
+ steps_out = ddim_timesteps + 1
59
+ if verbose:
60
+ print(f'Selected timesteps for ddim sampler: {steps_out}')
61
+ return steps_out
62
+
63
+
64
+ def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
65
+ # select alphas for computing the variance schedule
66
+ alphas = alphacums[ddim_timesteps]
67
+ alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
68
+
69
+ # according the the formula provided in https://arxiv.org/abs/2010.02502
70
+ sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
71
+ if verbose:
72
+ print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
73
+ print(f'For the chosen value of eta, which is {eta}, '
74
+ f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
75
+ return sigmas, alphas, alphas_prev
76
+
77
+
78
+ def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
79
+ """
80
+ Create a beta schedule that discretizes the given alpha_t_bar function,
81
+ which defines the cumulative product of (1-beta) over time from t = [0,1].
82
+ :param num_diffusion_timesteps: the number of betas to produce.
83
+ :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
84
+ produces the cumulative product of (1-beta) up to that
85
+ part of the diffusion process.
86
+ :param max_beta: the maximum beta to use; use values lower than 1 to
87
+ prevent singularities.
88
+ """
89
+ betas = []
90
+ for i in range(num_diffusion_timesteps):
91
+ t1 = i / num_diffusion_timesteps
92
+ t2 = (i + 1) / num_diffusion_timesteps
93
+ betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
94
+ return np.array(betas)
95
+
96
+
97
+ def extract_into_tensor(a, t, x_shape):
98
+ b, *_ = t.shape
99
+ out = a.gather(-1, t)
100
+ return out.reshape(b, *((1,) * (len(x_shape) - 1)))
101
+
102
+
103
+ def checkpoint(func, inputs, params, flag):
104
+ """
105
+ Evaluate a function without caching intermediate activations, allowing for
106
+ reduced memory at the expense of extra compute in the backward pass.
107
+ :param func: the function to evaluate.
108
+ :param inputs: the argument sequence to pass to `func`.
109
+ :param params: a sequence of parameters `func` depends on but does not
110
+ explicitly take as arguments.
111
+ :param flag: if False, disable gradient checkpointing.
112
+ """
113
+ if flag:
114
+ args = tuple(inputs) + tuple(params)
115
+ return CheckpointFunction.apply(func, len(inputs), *args)
116
+ else:
117
+ return func(*inputs)
118
+
119
+
120
+ class CheckpointFunction(torch.autograd.Function):
121
+ @staticmethod
122
+ def forward(ctx, run_function, length, *args):
123
+ ctx.run_function = run_function
124
+ ctx.input_tensors = list(args[:length])
125
+ ctx.input_params = list(args[length:])
126
+ ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
127
+ "dtype": torch.get_autocast_gpu_dtype(),
128
+ "cache_enabled": torch.is_autocast_cache_enabled()}
129
+ with torch.no_grad():
130
+ output_tensors = ctx.run_function(*ctx.input_tensors)
131
+ return output_tensors
132
+
133
+ @staticmethod
134
+ def backward(ctx, *output_grads):
135
+ ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
136
+ with torch.enable_grad(), \
137
+ torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
138
+ # Fixes a bug where the first op in run_function modifies the
139
+ # Tensor storage in place, which is not allowed for detach()'d
140
+ # Tensors.
141
+ shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
142
+ output_tensors = ctx.run_function(*shallow_copies)
143
+ input_grads = torch.autograd.grad(
144
+ output_tensors,
145
+ ctx.input_tensors + ctx.input_params,
146
+ output_grads,
147
+ allow_unused=True,
148
+ )
149
+ del ctx.input_tensors
150
+ del ctx.input_params
151
+ del output_tensors
152
+ return (None, None) + input_grads
153
+
154
+
155
+ def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
156
+ """
157
+ Create sinusoidal timestep embeddings.
158
+ :param timesteps: a 1-D Tensor of N indices, one per batch element.
159
+ These may be fractional.
160
+ :param dim: the dimension of the output.
161
+ :param max_period: controls the minimum frequency of the embeddings.
162
+ :return: an [N x dim] Tensor of positional embeddings.
163
+ """
164
+ if not repeat_only:
165
+ half = dim // 2
166
+ freqs = torch.exp(
167
+ -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
168
+ ).to(device=timesteps.device)
169
+ args = timesteps[:, None].float() * freqs[None]
170
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
171
+ if dim % 2:
172
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
173
+ else:
174
+ embedding = repeat(timesteps, 'b -> b d', d=dim)
175
+ return embedding
176
+
177
+
178
+ def zero_module(module):
179
+ """
180
+ Zero out the parameters of a module and return it.
181
+ """
182
+ for p in module.parameters():
183
+ p.detach().zero_()
184
+ return module
185
+
186
+
187
+ def scale_module(module, scale):
188
+ """
189
+ Scale the parameters of a module and return it.
190
+ """
191
+ for p in module.parameters():
192
+ p.detach().mul_(scale)
193
+ return module
194
+
195
+
196
+ def mean_flat(tensor):
197
+ """
198
+ Take the mean over all non-batch dimensions.
199
+ """
200
+ return tensor.mean(dim=list(range(1, len(tensor.shape))))
201
+
202
+
203
+ def normalization(channels):
204
+ """
205
+ Make a standard normalization layer.
206
+ :param channels: number of input channels.
207
+ :return: an nn.Module for normalization.
208
+ """
209
+ return GroupNorm32(32, channels)
210
+
211
+
212
+ # PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
213
+ class SiLU(nn.Module):
214
+ def forward(self, x):
215
+ return x * torch.sigmoid(x)
216
+
217
+
218
+ class GroupNorm32(nn.GroupNorm):
219
+ def forward(self, x):
220
+ return super().forward(x.float()).type(x.dtype)
221
+
222
+ def conv_nd(dims, *args, **kwargs):
223
+ """
224
+ Create a 1D, 2D, or 3D convolution module.
225
+ """
226
+ if dims == 1:
227
+ return nn.Conv1d(*args, **kwargs)
228
+ elif dims == 2:
229
+ return nn.Conv2d(*args, **kwargs)
230
+ elif dims == 3:
231
+ return nn.Conv3d(*args, **kwargs)
232
+ raise ValueError(f"unsupported dimensions: {dims}")
233
+
234
+
235
+ def linear(*args, **kwargs):
236
+ """
237
+ Create a linear module.
238
+ """
239
+ return nn.Linear(*args, **kwargs)
240
+
241
+
242
+ def avg_pool_nd(dims, *args, **kwargs):
243
+ """
244
+ Create a 1D, 2D, or 3D average pooling module.
245
+ """
246
+ if dims == 1:
247
+ return nn.AvgPool1d(*args, **kwargs)
248
+ elif dims == 2:
249
+ return nn.AvgPool2d(*args, **kwargs)
250
+ elif dims == 3:
251
+ return nn.AvgPool3d(*args, **kwargs)
252
+ raise ValueError(f"unsupported dimensions: {dims}")
253
+
254
+
255
+ class HybridConditioner(nn.Module):
256
+
257
+ def __init__(self, c_concat_config, c_crossattn_config):
258
+ super().__init__()
259
+ self.concat_conditioner = instantiate_from_config(c_concat_config)
260
+ self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
261
+
262
+ def forward(self, c_concat, c_crossattn):
263
+ c_concat = self.concat_conditioner(c_concat)
264
+ c_crossattn = self.crossattn_conditioner(c_crossattn)
265
+ return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
266
+
267
+
268
+ def noise_like(shape, device, repeat=False):
269
+ repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
270
+ noise = lambda: torch.randn(shape, device=device)
271
+ return repeat_noise() if repeat else noise()
ldm/modules/distributions/__init__.py ADDED
File without changes
ldm/modules/distributions/distributions.py ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import numpy as np
3
+
4
+
5
+ class AbstractDistribution:
6
+ def sample(self):
7
+ raise NotImplementedError()
8
+
9
+ def mode(self):
10
+ raise NotImplementedError()
11
+
12
+
13
+ class DiracDistribution(AbstractDistribution):
14
+ def __init__(self, value):
15
+ self.value = value
16
+
17
+ def sample(self):
18
+ return self.value
19
+
20
+ def mode(self):
21
+ return self.value
22
+
23
+
24
+ class DiagonalGaussianDistribution(object):
25
+ def __init__(self, parameters, deterministic=False):
26
+ self.parameters = parameters
27
+ self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
28
+ self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
29
+ self.deterministic = deterministic
30
+ self.std = torch.exp(0.5 * self.logvar)
31
+ self.var = torch.exp(self.logvar)
32
+ if self.deterministic:
33
+ self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
34
+
35
+ def sample(self):
36
+ x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
37
+ return x
38
+
39
+ def kl(self, other=None):
40
+ if self.deterministic:
41
+ return torch.Tensor([0.])
42
+ else:
43
+ if other is None:
44
+ return 0.5 * torch.sum(torch.pow(self.mean, 2)
45
+ + self.var - 1.0 - self.logvar,
46
+ dim=[1, 2, 3])
47
+ else:
48
+ return 0.5 * torch.sum(
49
+ torch.pow(self.mean - other.mean, 2) / other.var
50
+ + self.var / other.var - 1.0 - self.logvar + other.logvar,
51
+ dim=[1, 2, 3])
52
+
53
+ def nll(self, sample, dims=[1,2,3]):
54
+ if self.deterministic:
55
+ return torch.Tensor([0.])
56
+ logtwopi = np.log(2.0 * np.pi)
57
+ return 0.5 * torch.sum(
58
+ logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
59
+ dim=dims)
60
+
61
+ def mode(self):
62
+ return self.mean
63
+
64
+
65
+ def normal_kl(mean1, logvar1, mean2, logvar2):
66
+ """
67
+ source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
68
+ Compute the KL divergence between two gaussians.
69
+ Shapes are automatically broadcasted, so batches can be compared to
70
+ scalars, among other use cases.
71
+ """
72
+ tensor = None
73
+ for obj in (mean1, logvar1, mean2, logvar2):
74
+ if isinstance(obj, torch.Tensor):
75
+ tensor = obj
76
+ break
77
+ assert tensor is not None, "at least one argument must be a Tensor"
78
+
79
+ # Force variances to be Tensors. Broadcasting helps convert scalars to
80
+ # Tensors, but it does not work for torch.exp().
81
+ logvar1, logvar2 = [
82
+ x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
83
+ for x in (logvar1, logvar2)
84
+ ]
85
+
86
+ return 0.5 * (
87
+ -1.0
88
+ + logvar2
89
+ - logvar1
90
+ + torch.exp(logvar1 - logvar2)
91
+ + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
92
+ )
ldm/modules/ema.py ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+
4
+
5
+ class LitEma(nn.Module):
6
+ def __init__(self, model, decay=0.9999, use_num_upates=True):
7
+ super().__init__()
8
+ if decay < 0.0 or decay > 1.0:
9
+ raise ValueError('Decay must be between 0 and 1')
10
+
11
+ self.m_name2s_name = {}
12
+ self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
13
+ self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates
14
+ else torch.tensor(-1, dtype=torch.int))
15
+
16
+ for name, p in model.named_parameters():
17
+ if p.requires_grad:
18
+ # remove as '.'-character is not allowed in buffers
19
+ s_name = name.replace('.', '')
20
+ self.m_name2s_name.update({name: s_name})
21
+ self.register_buffer(s_name, p.clone().detach().data)
22
+
23
+ self.collected_params = []
24
+
25
+ def reset_num_updates(self):
26
+ del self.num_updates
27
+ self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int))
28
+
29
+ def forward(self, model):
30
+ decay = self.decay
31
+
32
+ if self.num_updates >= 0:
33
+ self.num_updates += 1
34
+ decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))
35
+
36
+ one_minus_decay = 1.0 - decay
37
+
38
+ with torch.no_grad():
39
+ m_param = dict(model.named_parameters())
40
+ shadow_params = dict(self.named_buffers())
41
+
42
+ for key in m_param:
43
+ if m_param[key].requires_grad:
44
+ sname = self.m_name2s_name[key]
45
+ shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
46
+ shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
47
+ else:
48
+ assert not key in self.m_name2s_name
49
+
50
+ def copy_to(self, model):
51
+ m_param = dict(model.named_parameters())
52
+ shadow_params = dict(self.named_buffers())
53
+ for key in m_param:
54
+ if m_param[key].requires_grad:
55
+ m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
56
+ else:
57
+ assert not key in self.m_name2s_name
58
+
59
+ def store(self, parameters):
60
+ """
61
+ Save the current parameters for restoring later.
62
+ Args:
63
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
64
+ temporarily stored.
65
+ """
66
+ self.collected_params = [param.clone() for param in parameters]
67
+
68
+ def restore(self, parameters):
69
+ """
70
+ Restore the parameters stored with the `store` method.
71
+ Useful to validate the model with EMA parameters without affecting the
72
+ original optimization process. Store the parameters before the
73
+ `copy_to` method. After validation (or model saving), use this to
74
+ restore the former parameters.
75
+ Args:
76
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
77
+ updated with the stored parameters.
78
+ """
79
+ for c_param, param in zip(self.collected_params, parameters):
80
+ param.data.copy_(c_param.data)
ldm/modules/encoders/__init__.py ADDED
File without changes
ldm/modules/encoders/modules.py ADDED
@@ -0,0 +1,213 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from torch.utils.checkpoint import checkpoint
4
+
5
+ from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
6
+
7
+ import open_clip
8
+ from ldm.util import default, count_params
9
+
10
+
11
+ class AbstractEncoder(nn.Module):
12
+ def __init__(self):
13
+ super().__init__()
14
+
15
+ def encode(self, *args, **kwargs):
16
+ raise NotImplementedError
17
+
18
+
19
+ class IdentityEncoder(AbstractEncoder):
20
+
21
+ def encode(self, x):
22
+ return x
23
+
24
+
25
+ class ClassEmbedder(nn.Module):
26
+ def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
27
+ super().__init__()
28
+ self.key = key
29
+ self.embedding = nn.Embedding(n_classes, embed_dim)
30
+ self.n_classes = n_classes
31
+ self.ucg_rate = ucg_rate
32
+
33
+ def forward(self, batch, key=None, disable_dropout=False):
34
+ if key is None:
35
+ key = self.key
36
+ # this is for use in crossattn
37
+ c = batch[key][:, None]
38
+ if self.ucg_rate > 0. and not disable_dropout:
39
+ mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
40
+ c = mask * c + (1-mask) * torch.ones_like(c)*(self.n_classes-1)
41
+ c = c.long()
42
+ c = self.embedding(c)
43
+ return c
44
+
45
+ def get_unconditional_conditioning(self, bs, device="cuda"):
46
+ uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
47
+ uc = torch.ones((bs,), device=device) * uc_class
48
+ uc = {self.key: uc}
49
+ return uc
50
+
51
+
52
+ def disabled_train(self, mode=True):
53
+ """Overwrite model.train with this function to make sure train/eval mode
54
+ does not change anymore."""
55
+ return self
56
+
57
+
58
+ class FrozenT5Embedder(AbstractEncoder):
59
+ """Uses the T5 transformer encoder for text"""
60
+ def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
61
+ super().__init__()
62
+ self.tokenizer = T5Tokenizer.from_pretrained(version)
63
+ self.transformer = T5EncoderModel.from_pretrained(version)
64
+ self.device = device
65
+ self.max_length = max_length # TODO: typical value?
66
+ if freeze:
67
+ self.freeze()
68
+
69
+ def freeze(self):
70
+ self.transformer = self.transformer.eval()
71
+ #self.train = disabled_train
72
+ for param in self.parameters():
73
+ param.requires_grad = False
74
+
75
+ def forward(self, text):
76
+ batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
77
+ return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
78
+ tokens = batch_encoding["input_ids"].to(self.device)
79
+ outputs = self.transformer(input_ids=tokens)
80
+
81
+ z = outputs.last_hidden_state
82
+ return z
83
+
84
+ def encode(self, text):
85
+ return self(text)
86
+
87
+
88
+ class FrozenCLIPEmbedder(AbstractEncoder):
89
+ """Uses the CLIP transformer encoder for text (from huggingface)"""
90
+ LAYERS = [
91
+ "last",
92
+ "pooled",
93
+ "hidden"
94
+ ]
95
+ def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
96
+ freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32
97
+ super().__init__()
98
+ assert layer in self.LAYERS
99
+ self.tokenizer = CLIPTokenizer.from_pretrained(version)
100
+ self.transformer = CLIPTextModel.from_pretrained(version)
101
+ self.device = device
102
+ self.max_length = max_length
103
+ if freeze:
104
+ self.freeze()
105
+ self.layer = layer
106
+ self.layer_idx = layer_idx
107
+ if layer == "hidden":
108
+ assert layer_idx is not None
109
+ assert 0 <= abs(layer_idx) <= 12
110
+
111
+ def freeze(self):
112
+ self.transformer = self.transformer.eval()
113
+ #self.train = disabled_train
114
+ for param in self.parameters():
115
+ param.requires_grad = False
116
+
117
+ def forward(self, text):
118
+ batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
119
+ return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
120
+ tokens = batch_encoding["input_ids"].to(self.device)
121
+ outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden")
122
+ if self.layer == "last":
123
+ z = outputs.last_hidden_state
124
+ elif self.layer == "pooled":
125
+ z = outputs.pooler_output[:, None, :]
126
+ else:
127
+ z = outputs.hidden_states[self.layer_idx]
128
+ return z
129
+
130
+ def encode(self, text):
131
+ return self(text)
132
+
133
+
134
+ class FrozenOpenCLIPEmbedder(AbstractEncoder):
135
+ """
136
+ Uses the OpenCLIP transformer encoder for text
137
+ """
138
+ LAYERS = [
139
+ #"pooled",
140
+ "last",
141
+ "penultimate"
142
+ ]
143
+ def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
144
+ freeze=True, layer="last"):
145
+ super().__init__()
146
+ assert layer in self.LAYERS
147
+ model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version)
148
+ del model.visual
149
+ self.model = model
150
+
151
+ self.device = device
152
+ self.max_length = max_length
153
+ if freeze:
154
+ self.freeze()
155
+ self.layer = layer
156
+ if self.layer == "last":
157
+ self.layer_idx = 0
158
+ elif self.layer == "penultimate":
159
+ self.layer_idx = 1
160
+ else:
161
+ raise NotImplementedError()
162
+
163
+ def freeze(self):
164
+ self.model = self.model.eval()
165
+ for param in self.parameters():
166
+ param.requires_grad = False
167
+
168
+ def forward(self, text):
169
+ tokens = open_clip.tokenize(text)
170
+ z = self.encode_with_transformer(tokens.to(self.device))
171
+ return z
172
+
173
+ def encode_with_transformer(self, text):
174
+ x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model]
175
+ x = x + self.model.positional_embedding
176
+ x = x.permute(1, 0, 2) # NLD -> LND
177
+ x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
178
+ x = x.permute(1, 0, 2) # LND -> NLD
179
+ x = self.model.ln_final(x)
180
+ return x
181
+
182
+ def text_transformer_forward(self, x: torch.Tensor, attn_mask = None):
183
+ for i, r in enumerate(self.model.transformer.resblocks):
184
+ if i == len(self.model.transformer.resblocks) - self.layer_idx:
185
+ break
186
+ if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
187
+ x = checkpoint(r, x, attn_mask)
188
+ else:
189
+ x = r(x, attn_mask=attn_mask)
190
+ return x
191
+
192
+ def encode(self, text):
193
+ return self(text)
194
+
195
+
196
+ class FrozenCLIPT5Encoder(AbstractEncoder):
197
+ def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
198
+ clip_max_length=77, t5_max_length=77):
199
+ super().__init__()
200
+ self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
201
+ self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
202
+ print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder)*1.e-6:.2f} M parameters, "
203
+ f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder)*1.e-6:.2f} M params.")
204
+
205
+ def encode(self, text):
206
+ return self(text)
207
+
208
+ def forward(self, text):
209
+ clip_z = self.clip_encoder.encode(text)
210
+ t5_z = self.t5_encoder.encode(text)
211
+ return [clip_z, t5_z]
212
+
213
+
ldm/modules/image_degradation/__init__.py ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr
2
+ from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light
ldm/modules/image_degradation/bsrgan.py ADDED
@@ -0,0 +1,730 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ """
3
+ # --------------------------------------------
4
+ # Super-Resolution
5
+ # --------------------------------------------
6
+ #
7
+ # Kai Zhang ([email protected])
8
+ # https://github.com/cszn
9
+ # From 2019/03--2021/08
10
+ # --------------------------------------------
11
+ """
12
+
13
+ import numpy as np
14
+ import cv2
15
+ import torch
16
+
17
+ from functools import partial
18
+ import random
19
+ from scipy import ndimage
20
+ import scipy
21
+ import scipy.stats as ss
22
+ from scipy.interpolate import interp2d
23
+ from scipy.linalg import orth
24
+ import albumentations
25
+
26
+ import ldm.modules.image_degradation.utils_image as util
27
+
28
+
29
+ def modcrop_np(img, sf):
30
+ '''
31
+ Args:
32
+ img: numpy image, WxH or WxHxC
33
+ sf: scale factor
34
+ Return:
35
+ cropped image
36
+ '''
37
+ w, h = img.shape[:2]
38
+ im = np.copy(img)
39
+ return im[:w - w % sf, :h - h % sf, ...]
40
+
41
+
42
+ """
43
+ # --------------------------------------------
44
+ # anisotropic Gaussian kernels
45
+ # --------------------------------------------
46
+ """
47
+
48
+
49
+ def analytic_kernel(k):
50
+ """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
51
+ k_size = k.shape[0]
52
+ # Calculate the big kernels size
53
+ big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
54
+ # Loop over the small kernel to fill the big one
55
+ for r in range(k_size):
56
+ for c in range(k_size):
57
+ big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
58
+ # Crop the edges of the big kernel to ignore very small values and increase run time of SR
59
+ crop = k_size // 2
60
+ cropped_big_k = big_k[crop:-crop, crop:-crop]
61
+ # Normalize to 1
62
+ return cropped_big_k / cropped_big_k.sum()
63
+
64
+
65
+ def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
66
+ """ generate an anisotropic Gaussian kernel
67
+ Args:
68
+ ksize : e.g., 15, kernel size
69
+ theta : [0, pi], rotation angle range
70
+ l1 : [0.1,50], scaling of eigenvalues
71
+ l2 : [0.1,l1], scaling of eigenvalues
72
+ If l1 = l2, will get an isotropic Gaussian kernel.
73
+ Returns:
74
+ k : kernel
75
+ """
76
+
77
+ v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
78
+ V = np.array([[v[0], v[1]], [v[1], -v[0]]])
79
+ D = np.array([[l1, 0], [0, l2]])
80
+ Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
81
+ k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
82
+
83
+ return k
84
+
85
+
86
+ def gm_blur_kernel(mean, cov, size=15):
87
+ center = size / 2.0 + 0.5
88
+ k = np.zeros([size, size])
89
+ for y in range(size):
90
+ for x in range(size):
91
+ cy = y - center + 1
92
+ cx = x - center + 1
93
+ k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
94
+
95
+ k = k / np.sum(k)
96
+ return k
97
+
98
+
99
+ def shift_pixel(x, sf, upper_left=True):
100
+ """shift pixel for super-resolution with different scale factors
101
+ Args:
102
+ x: WxHxC or WxH
103
+ sf: scale factor
104
+ upper_left: shift direction
105
+ """
106
+ h, w = x.shape[:2]
107
+ shift = (sf - 1) * 0.5
108
+ xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
109
+ if upper_left:
110
+ x1 = xv + shift
111
+ y1 = yv + shift
112
+ else:
113
+ x1 = xv - shift
114
+ y1 = yv - shift
115
+
116
+ x1 = np.clip(x1, 0, w - 1)
117
+ y1 = np.clip(y1, 0, h - 1)
118
+
119
+ if x.ndim == 2:
120
+ x = interp2d(xv, yv, x)(x1, y1)
121
+ if x.ndim == 3:
122
+ for i in range(x.shape[-1]):
123
+ x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
124
+
125
+ return x
126
+
127
+
128
+ def blur(x, k):
129
+ '''
130
+ x: image, NxcxHxW
131
+ k: kernel, Nx1xhxw
132
+ '''
133
+ n, c = x.shape[:2]
134
+ p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
135
+ x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
136
+ k = k.repeat(1, c, 1, 1)
137
+ k = k.view(-1, 1, k.shape[2], k.shape[3])
138
+ x = x.view(1, -1, x.shape[2], x.shape[3])
139
+ x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
140
+ x = x.view(n, c, x.shape[2], x.shape[3])
141
+
142
+ return x
143
+
144
+
145
+ def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
146
+ """"
147
+ # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
148
+ # Kai Zhang
149
+ # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var
150
+ # max_var = 2.5 * sf
151
+ """
152
+ # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
153
+ lambda_1 = min_var + np.random.rand() * (max_var - min_var)
154
+ lambda_2 = min_var + np.random.rand() * (max_var - min_var)
155
+ theta = np.random.rand() * np.pi # random theta
156
+ noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
157
+
158
+ # Set COV matrix using Lambdas and Theta
159
+ LAMBDA = np.diag([lambda_1, lambda_2])
160
+ Q = np.array([[np.cos(theta), -np.sin(theta)],
161
+ [np.sin(theta), np.cos(theta)]])
162
+ SIGMA = Q @ LAMBDA @ Q.T
163
+ INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
164
+
165
+ # Set expectation position (shifting kernel for aligned image)
166
+ MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2)
167
+ MU = MU[None, None, :, None]
168
+
169
+ # Create meshgrid for Gaussian
170
+ [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
171
+ Z = np.stack([X, Y], 2)[:, :, :, None]
172
+
173
+ # Calcualte Gaussian for every pixel of the kernel
174
+ ZZ = Z - MU
175
+ ZZ_t = ZZ.transpose(0, 1, 3, 2)
176
+ raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
177
+
178
+ # shift the kernel so it will be centered
179
+ # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
180
+
181
+ # Normalize the kernel and return
182
+ # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
183
+ kernel = raw_kernel / np.sum(raw_kernel)
184
+ return kernel
185
+
186
+
187
+ def fspecial_gaussian(hsize, sigma):
188
+ hsize = [hsize, hsize]
189
+ siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
190
+ std = sigma
191
+ [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
192
+ arg = -(x * x + y * y) / (2 * std * std)
193
+ h = np.exp(arg)
194
+ h[h < scipy.finfo(float).eps * h.max()] = 0
195
+ sumh = h.sum()
196
+ if sumh != 0:
197
+ h = h / sumh
198
+ return h
199
+
200
+
201
+ def fspecial_laplacian(alpha):
202
+ alpha = max([0, min([alpha, 1])])
203
+ h1 = alpha / (alpha + 1)
204
+ h2 = (1 - alpha) / (alpha + 1)
205
+ h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
206
+ h = np.array(h)
207
+ return h
208
+
209
+
210
+ def fspecial(filter_type, *args, **kwargs):
211
+ '''
212
+ python code from:
213
+ https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
214
+ '''
215
+ if filter_type == 'gaussian':
216
+ return fspecial_gaussian(*args, **kwargs)
217
+ if filter_type == 'laplacian':
218
+ return fspecial_laplacian(*args, **kwargs)
219
+
220
+
221
+ """
222
+ # --------------------------------------------
223
+ # degradation models
224
+ # --------------------------------------------
225
+ """
226
+
227
+
228
+ def bicubic_degradation(x, sf=3):
229
+ '''
230
+ Args:
231
+ x: HxWxC image, [0, 1]
232
+ sf: down-scale factor
233
+ Return:
234
+ bicubicly downsampled LR image
235
+ '''
236
+ x = util.imresize_np(x, scale=1 / sf)
237
+ return x
238
+
239
+
240
+ def srmd_degradation(x, k, sf=3):
241
+ ''' blur + bicubic downsampling
242
+ Args:
243
+ x: HxWxC image, [0, 1]
244
+ k: hxw, double
245
+ sf: down-scale factor
246
+ Return:
247
+ downsampled LR image
248
+ Reference:
249
+ @inproceedings{zhang2018learning,
250
+ title={Learning a single convolutional super-resolution network for multiple degradations},
251
+ author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
252
+ booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
253
+ pages={3262--3271},
254
+ year={2018}
255
+ }
256
+ '''
257
+ x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror'
258
+ x = bicubic_degradation(x, sf=sf)
259
+ return x
260
+
261
+
262
+ def dpsr_degradation(x, k, sf=3):
263
+ ''' bicubic downsampling + blur
264
+ Args:
265
+ x: HxWxC image, [0, 1]
266
+ k: hxw, double
267
+ sf: down-scale factor
268
+ Return:
269
+ downsampled LR image
270
+ Reference:
271
+ @inproceedings{zhang2019deep,
272
+ title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
273
+ author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
274
+ booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
275
+ pages={1671--1681},
276
+ year={2019}
277
+ }
278
+ '''
279
+ x = bicubic_degradation(x, sf=sf)
280
+ x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
281
+ return x
282
+
283
+
284
+ def classical_degradation(x, k, sf=3):
285
+ ''' blur + downsampling
286
+ Args:
287
+ x: HxWxC image, [0, 1]/[0, 255]
288
+ k: hxw, double
289
+ sf: down-scale factor
290
+ Return:
291
+ downsampled LR image
292
+ '''
293
+ x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
294
+ # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
295
+ st = 0
296
+ return x[st::sf, st::sf, ...]
297
+
298
+
299
+ def add_sharpening(img, weight=0.5, radius=50, threshold=10):
300
+ """USM sharpening. borrowed from real-ESRGAN
301
+ Input image: I; Blurry image: B.
302
+ 1. K = I + weight * (I - B)
303
+ 2. Mask = 1 if abs(I - B) > threshold, else: 0
304
+ 3. Blur mask:
305
+ 4. Out = Mask * K + (1 - Mask) * I
306
+ Args:
307
+ img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
308
+ weight (float): Sharp weight. Default: 1.
309
+ radius (float): Kernel size of Gaussian blur. Default: 50.
310
+ threshold (int):
311
+ """
312
+ if radius % 2 == 0:
313
+ radius += 1
314
+ blur = cv2.GaussianBlur(img, (radius, radius), 0)
315
+ residual = img - blur
316
+ mask = np.abs(residual) * 255 > threshold
317
+ mask = mask.astype('float32')
318
+ soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
319
+
320
+ K = img + weight * residual
321
+ K = np.clip(K, 0, 1)
322
+ return soft_mask * K + (1 - soft_mask) * img
323
+
324
+
325
+ def add_blur(img, sf=4):
326
+ wd2 = 4.0 + sf
327
+ wd = 2.0 + 0.2 * sf
328
+ if random.random() < 0.5:
329
+ l1 = wd2 * random.random()
330
+ l2 = wd2 * random.random()
331
+ k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
332
+ else:
333
+ k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random())
334
+ img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
335
+
336
+ return img
337
+
338
+
339
+ def add_resize(img, sf=4):
340
+ rnum = np.random.rand()
341
+ if rnum > 0.8: # up
342
+ sf1 = random.uniform(1, 2)
343
+ elif rnum < 0.7: # down
344
+ sf1 = random.uniform(0.5 / sf, 1)
345
+ else:
346
+ sf1 = 1.0
347
+ img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
348
+ img = np.clip(img, 0.0, 1.0)
349
+
350
+ return img
351
+
352
+
353
+ # def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
354
+ # noise_level = random.randint(noise_level1, noise_level2)
355
+ # rnum = np.random.rand()
356
+ # if rnum > 0.6: # add color Gaussian noise
357
+ # img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
358
+ # elif rnum < 0.4: # add grayscale Gaussian noise
359
+ # img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
360
+ # else: # add noise
361
+ # L = noise_level2 / 255.
362
+ # D = np.diag(np.random.rand(3))
363
+ # U = orth(np.random.rand(3, 3))
364
+ # conv = np.dot(np.dot(np.transpose(U), D), U)
365
+ # img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
366
+ # img = np.clip(img, 0.0, 1.0)
367
+ # return img
368
+
369
+ def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
370
+ noise_level = random.randint(noise_level1, noise_level2)
371
+ rnum = np.random.rand()
372
+ if rnum > 0.6: # add color Gaussian noise
373
+ img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
374
+ elif rnum < 0.4: # add grayscale Gaussian noise
375
+ img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
376
+ else: # add noise
377
+ L = noise_level2 / 255.
378
+ D = np.diag(np.random.rand(3))
379
+ U = orth(np.random.rand(3, 3))
380
+ conv = np.dot(np.dot(np.transpose(U), D), U)
381
+ img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
382
+ img = np.clip(img, 0.0, 1.0)
383
+ return img
384
+
385
+
386
+ def add_speckle_noise(img, noise_level1=2, noise_level2=25):
387
+ noise_level = random.randint(noise_level1, noise_level2)
388
+ img = np.clip(img, 0.0, 1.0)
389
+ rnum = random.random()
390
+ if rnum > 0.6:
391
+ img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
392
+ elif rnum < 0.4:
393
+ img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
394
+ else:
395
+ L = noise_level2 / 255.
396
+ D = np.diag(np.random.rand(3))
397
+ U = orth(np.random.rand(3, 3))
398
+ conv = np.dot(np.dot(np.transpose(U), D), U)
399
+ img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
400
+ img = np.clip(img, 0.0, 1.0)
401
+ return img
402
+
403
+
404
+ def add_Poisson_noise(img):
405
+ img = np.clip((img * 255.0).round(), 0, 255) / 255.
406
+ vals = 10 ** (2 * random.random() + 2.0) # [2, 4]
407
+ if random.random() < 0.5:
408
+ img = np.random.poisson(img * vals).astype(np.float32) / vals
409
+ else:
410
+ img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
411
+ img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
412
+ noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
413
+ img += noise_gray[:, :, np.newaxis]
414
+ img = np.clip(img, 0.0, 1.0)
415
+ return img
416
+
417
+
418
+ def add_JPEG_noise(img):
419
+ quality_factor = random.randint(30, 95)
420
+ img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
421
+ result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
422
+ img = cv2.imdecode(encimg, 1)
423
+ img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
424
+ return img
425
+
426
+
427
+ def random_crop(lq, hq, sf=4, lq_patchsize=64):
428
+ h, w = lq.shape[:2]
429
+ rnd_h = random.randint(0, h - lq_patchsize)
430
+ rnd_w = random.randint(0, w - lq_patchsize)
431
+ lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
432
+
433
+ rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
434
+ hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
435
+ return lq, hq
436
+
437
+
438
+ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
439
+ """
440
+ This is the degradation model of BSRGAN from the paper
441
+ "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
442
+ ----------
443
+ img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
444
+ sf: scale factor
445
+ isp_model: camera ISP model
446
+ Returns
447
+ -------
448
+ img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
449
+ hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
450
+ """
451
+ isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
452
+ sf_ori = sf
453
+
454
+ h1, w1 = img.shape[:2]
455
+ img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
456
+ h, w = img.shape[:2]
457
+
458
+ if h < lq_patchsize * sf or w < lq_patchsize * sf:
459
+ raise ValueError(f'img size ({h1}X{w1}) is too small!')
460
+
461
+ hq = img.copy()
462
+
463
+ if sf == 4 and random.random() < scale2_prob: # downsample1
464
+ if np.random.rand() < 0.5:
465
+ img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
466
+ interpolation=random.choice([1, 2, 3]))
467
+ else:
468
+ img = util.imresize_np(img, 1 / 2, True)
469
+ img = np.clip(img, 0.0, 1.0)
470
+ sf = 2
471
+
472
+ shuffle_order = random.sample(range(7), 7)
473
+ idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
474
+ if idx1 > idx2: # keep downsample3 last
475
+ shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
476
+
477
+ for i in shuffle_order:
478
+
479
+ if i == 0:
480
+ img = add_blur(img, sf=sf)
481
+
482
+ elif i == 1:
483
+ img = add_blur(img, sf=sf)
484
+
485
+ elif i == 2:
486
+ a, b = img.shape[1], img.shape[0]
487
+ # downsample2
488
+ if random.random() < 0.75:
489
+ sf1 = random.uniform(1, 2 * sf)
490
+ img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
491
+ interpolation=random.choice([1, 2, 3]))
492
+ else:
493
+ k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
494
+ k_shifted = shift_pixel(k, sf)
495
+ k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
496
+ img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
497
+ img = img[0::sf, 0::sf, ...] # nearest downsampling
498
+ img = np.clip(img, 0.0, 1.0)
499
+
500
+ elif i == 3:
501
+ # downsample3
502
+ img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
503
+ img = np.clip(img, 0.0, 1.0)
504
+
505
+ elif i == 4:
506
+ # add Gaussian noise
507
+ img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
508
+
509
+ elif i == 5:
510
+ # add JPEG noise
511
+ if random.random() < jpeg_prob:
512
+ img = add_JPEG_noise(img)
513
+
514
+ elif i == 6:
515
+ # add processed camera sensor noise
516
+ if random.random() < isp_prob and isp_model is not None:
517
+ with torch.no_grad():
518
+ img, hq = isp_model.forward(img.copy(), hq)
519
+
520
+ # add final JPEG compression noise
521
+ img = add_JPEG_noise(img)
522
+
523
+ # random crop
524
+ img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
525
+
526
+ return img, hq
527
+
528
+
529
+ # todo no isp_model?
530
+ def degradation_bsrgan_variant(image, sf=4, isp_model=None):
531
+ """
532
+ This is the degradation model of BSRGAN from the paper
533
+ "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
534
+ ----------
535
+ sf: scale factor
536
+ isp_model: camera ISP model
537
+ Returns
538
+ -------
539
+ img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
540
+ hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
541
+ """
542
+ image = util.uint2single(image)
543
+ isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
544
+ sf_ori = sf
545
+
546
+ h1, w1 = image.shape[:2]
547
+ image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
548
+ h, w = image.shape[:2]
549
+
550
+ hq = image.copy()
551
+
552
+ if sf == 4 and random.random() < scale2_prob: # downsample1
553
+ if np.random.rand() < 0.5:
554
+ image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
555
+ interpolation=random.choice([1, 2, 3]))
556
+ else:
557
+ image = util.imresize_np(image, 1 / 2, True)
558
+ image = np.clip(image, 0.0, 1.0)
559
+ sf = 2
560
+
561
+ shuffle_order = random.sample(range(7), 7)
562
+ idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
563
+ if idx1 > idx2: # keep downsample3 last
564
+ shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
565
+
566
+ for i in shuffle_order:
567
+
568
+ if i == 0:
569
+ image = add_blur(image, sf=sf)
570
+
571
+ elif i == 1:
572
+ image = add_blur(image, sf=sf)
573
+
574
+ elif i == 2:
575
+ a, b = image.shape[1], image.shape[0]
576
+ # downsample2
577
+ if random.random() < 0.75:
578
+ sf1 = random.uniform(1, 2 * sf)
579
+ image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
580
+ interpolation=random.choice([1, 2, 3]))
581
+ else:
582
+ k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
583
+ k_shifted = shift_pixel(k, sf)
584
+ k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
585
+ image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
586
+ image = image[0::sf, 0::sf, ...] # nearest downsampling
587
+ image = np.clip(image, 0.0, 1.0)
588
+
589
+ elif i == 3:
590
+ # downsample3
591
+ image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
592
+ image = np.clip(image, 0.0, 1.0)
593
+
594
+ elif i == 4:
595
+ # add Gaussian noise
596
+ image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25)
597
+
598
+ elif i == 5:
599
+ # add JPEG noise
600
+ if random.random() < jpeg_prob:
601
+ image = add_JPEG_noise(image)
602
+
603
+ # elif i == 6:
604
+ # # add processed camera sensor noise
605
+ # if random.random() < isp_prob and isp_model is not None:
606
+ # with torch.no_grad():
607
+ # img, hq = isp_model.forward(img.copy(), hq)
608
+
609
+ # add final JPEG compression noise
610
+ image = add_JPEG_noise(image)
611
+ image = util.single2uint(image)
612
+ example = {"image":image}
613
+ return example
614
+
615
+
616
+ # TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc...
617
+ def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None):
618
+ """
619
+ This is an extended degradation model by combining
620
+ the degradation models of BSRGAN and Real-ESRGAN
621
+ ----------
622
+ img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
623
+ sf: scale factor
624
+ use_shuffle: the degradation shuffle
625
+ use_sharp: sharpening the img
626
+ Returns
627
+ -------
628
+ img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
629
+ hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
630
+ """
631
+
632
+ h1, w1 = img.shape[:2]
633
+ img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
634
+ h, w = img.shape[:2]
635
+
636
+ if h < lq_patchsize * sf or w < lq_patchsize * sf:
637
+ raise ValueError(f'img size ({h1}X{w1}) is too small!')
638
+
639
+ if use_sharp:
640
+ img = add_sharpening(img)
641
+ hq = img.copy()
642
+
643
+ if random.random() < shuffle_prob:
644
+ shuffle_order = random.sample(range(13), 13)
645
+ else:
646
+ shuffle_order = list(range(13))
647
+ # local shuffle for noise, JPEG is always the last one
648
+ shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6)))
649
+ shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13)))
650
+
651
+ poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1
652
+
653
+ for i in shuffle_order:
654
+ if i == 0:
655
+ img = add_blur(img, sf=sf)
656
+ elif i == 1:
657
+ img = add_resize(img, sf=sf)
658
+ elif i == 2:
659
+ img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
660
+ elif i == 3:
661
+ if random.random() < poisson_prob:
662
+ img = add_Poisson_noise(img)
663
+ elif i == 4:
664
+ if random.random() < speckle_prob:
665
+ img = add_speckle_noise(img)
666
+ elif i == 5:
667
+ if random.random() < isp_prob and isp_model is not None:
668
+ with torch.no_grad():
669
+ img, hq = isp_model.forward(img.copy(), hq)
670
+ elif i == 6:
671
+ img = add_JPEG_noise(img)
672
+ elif i == 7:
673
+ img = add_blur(img, sf=sf)
674
+ elif i == 8:
675
+ img = add_resize(img, sf=sf)
676
+ elif i == 9:
677
+ img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
678
+ elif i == 10:
679
+ if random.random() < poisson_prob:
680
+ img = add_Poisson_noise(img)
681
+ elif i == 11:
682
+ if random.random() < speckle_prob:
683
+ img = add_speckle_noise(img)
684
+ elif i == 12:
685
+ if random.random() < isp_prob and isp_model is not None:
686
+ with torch.no_grad():
687
+ img, hq = isp_model.forward(img.copy(), hq)
688
+ else:
689
+ print('check the shuffle!')
690
+
691
+ # resize to desired size
692
+ img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])),
693
+ interpolation=random.choice([1, 2, 3]))
694
+
695
+ # add final JPEG compression noise
696
+ img = add_JPEG_noise(img)
697
+
698
+ # random crop
699
+ img, hq = random_crop(img, hq, sf, lq_patchsize)
700
+
701
+ return img, hq
702
+
703
+
704
+ if __name__ == '__main__':
705
+ print("hey")
706
+ img = util.imread_uint('utils/test.png', 3)
707
+ print(img)
708
+ img = util.uint2single(img)
709
+ print(img)
710
+ img = img[:448, :448]
711
+ h = img.shape[0] // 4
712
+ print("resizing to", h)
713
+ sf = 4
714
+ deg_fn = partial(degradation_bsrgan_variant, sf=sf)
715
+ for i in range(20):
716
+ print(i)
717
+ img_lq = deg_fn(img)
718
+ print(img_lq)
719
+ img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"]
720
+ print(img_lq.shape)
721
+ print("bicubic", img_lq_bicubic.shape)
722
+ print(img_hq.shape)
723
+ lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
724
+ interpolation=0)
725
+ lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
726
+ interpolation=0)
727
+ img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
728
+ util.imsave(img_concat, str(i) + '.png')
729
+
730
+
ldm/modules/image_degradation/bsrgan_light.py ADDED
@@ -0,0 +1,651 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ import numpy as np
3
+ import cv2
4
+ import torch
5
+
6
+ from functools import partial
7
+ import random
8
+ from scipy import ndimage
9
+ import scipy
10
+ import scipy.stats as ss
11
+ from scipy.interpolate import interp2d
12
+ from scipy.linalg import orth
13
+ import albumentations
14
+
15
+ import ldm.modules.image_degradation.utils_image as util
16
+
17
+ """
18
+ # --------------------------------------------
19
+ # Super-Resolution
20
+ # --------------------------------------------
21
+ #
22
+ # Kai Zhang ([email protected])
23
+ # https://github.com/cszn
24
+ # From 2019/03--2021/08
25
+ # --------------------------------------------
26
+ """
27
+
28
+ def modcrop_np(img, sf):
29
+ '''
30
+ Args:
31
+ img: numpy image, WxH or WxHxC
32
+ sf: scale factor
33
+ Return:
34
+ cropped image
35
+ '''
36
+ w, h = img.shape[:2]
37
+ im = np.copy(img)
38
+ return im[:w - w % sf, :h - h % sf, ...]
39
+
40
+
41
+ """
42
+ # --------------------------------------------
43
+ # anisotropic Gaussian kernels
44
+ # --------------------------------------------
45
+ """
46
+
47
+
48
+ def analytic_kernel(k):
49
+ """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
50
+ k_size = k.shape[0]
51
+ # Calculate the big kernels size
52
+ big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
53
+ # Loop over the small kernel to fill the big one
54
+ for r in range(k_size):
55
+ for c in range(k_size):
56
+ big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
57
+ # Crop the edges of the big kernel to ignore very small values and increase run time of SR
58
+ crop = k_size // 2
59
+ cropped_big_k = big_k[crop:-crop, crop:-crop]
60
+ # Normalize to 1
61
+ return cropped_big_k / cropped_big_k.sum()
62
+
63
+
64
+ def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
65
+ """ generate an anisotropic Gaussian kernel
66
+ Args:
67
+ ksize : e.g., 15, kernel size
68
+ theta : [0, pi], rotation angle range
69
+ l1 : [0.1,50], scaling of eigenvalues
70
+ l2 : [0.1,l1], scaling of eigenvalues
71
+ If l1 = l2, will get an isotropic Gaussian kernel.
72
+ Returns:
73
+ k : kernel
74
+ """
75
+
76
+ v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
77
+ V = np.array([[v[0], v[1]], [v[1], -v[0]]])
78
+ D = np.array([[l1, 0], [0, l2]])
79
+ Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
80
+ k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
81
+
82
+ return k
83
+
84
+
85
+ def gm_blur_kernel(mean, cov, size=15):
86
+ center = size / 2.0 + 0.5
87
+ k = np.zeros([size, size])
88
+ for y in range(size):
89
+ for x in range(size):
90
+ cy = y - center + 1
91
+ cx = x - center + 1
92
+ k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
93
+
94
+ k = k / np.sum(k)
95
+ return k
96
+
97
+
98
+ def shift_pixel(x, sf, upper_left=True):
99
+ """shift pixel for super-resolution with different scale factors
100
+ Args:
101
+ x: WxHxC or WxH
102
+ sf: scale factor
103
+ upper_left: shift direction
104
+ """
105
+ h, w = x.shape[:2]
106
+ shift = (sf - 1) * 0.5
107
+ xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
108
+ if upper_left:
109
+ x1 = xv + shift
110
+ y1 = yv + shift
111
+ else:
112
+ x1 = xv - shift
113
+ y1 = yv - shift
114
+
115
+ x1 = np.clip(x1, 0, w - 1)
116
+ y1 = np.clip(y1, 0, h - 1)
117
+
118
+ if x.ndim == 2:
119
+ x = interp2d(xv, yv, x)(x1, y1)
120
+ if x.ndim == 3:
121
+ for i in range(x.shape[-1]):
122
+ x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
123
+
124
+ return x
125
+
126
+
127
+ def blur(x, k):
128
+ '''
129
+ x: image, NxcxHxW
130
+ k: kernel, Nx1xhxw
131
+ '''
132
+ n, c = x.shape[:2]
133
+ p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
134
+ x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
135
+ k = k.repeat(1, c, 1, 1)
136
+ k = k.view(-1, 1, k.shape[2], k.shape[3])
137
+ x = x.view(1, -1, x.shape[2], x.shape[3])
138
+ x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
139
+ x = x.view(n, c, x.shape[2], x.shape[3])
140
+
141
+ return x
142
+
143
+
144
+ def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
145
+ """"
146
+ # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
147
+ # Kai Zhang
148
+ # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var
149
+ # max_var = 2.5 * sf
150
+ """
151
+ # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
152
+ lambda_1 = min_var + np.random.rand() * (max_var - min_var)
153
+ lambda_2 = min_var + np.random.rand() * (max_var - min_var)
154
+ theta = np.random.rand() * np.pi # random theta
155
+ noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
156
+
157
+ # Set COV matrix using Lambdas and Theta
158
+ LAMBDA = np.diag([lambda_1, lambda_2])
159
+ Q = np.array([[np.cos(theta), -np.sin(theta)],
160
+ [np.sin(theta), np.cos(theta)]])
161
+ SIGMA = Q @ LAMBDA @ Q.T
162
+ INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
163
+
164
+ # Set expectation position (shifting kernel for aligned image)
165
+ MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2)
166
+ MU = MU[None, None, :, None]
167
+
168
+ # Create meshgrid for Gaussian
169
+ [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
170
+ Z = np.stack([X, Y], 2)[:, :, :, None]
171
+
172
+ # Calcualte Gaussian for every pixel of the kernel
173
+ ZZ = Z - MU
174
+ ZZ_t = ZZ.transpose(0, 1, 3, 2)
175
+ raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
176
+
177
+ # shift the kernel so it will be centered
178
+ # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
179
+
180
+ # Normalize the kernel and return
181
+ # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
182
+ kernel = raw_kernel / np.sum(raw_kernel)
183
+ return kernel
184
+
185
+
186
+ def fspecial_gaussian(hsize, sigma):
187
+ hsize = [hsize, hsize]
188
+ siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
189
+ std = sigma
190
+ [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
191
+ arg = -(x * x + y * y) / (2 * std * std)
192
+ h = np.exp(arg)
193
+ h[h < scipy.finfo(float).eps * h.max()] = 0
194
+ sumh = h.sum()
195
+ if sumh != 0:
196
+ h = h / sumh
197
+ return h
198
+
199
+
200
+ def fspecial_laplacian(alpha):
201
+ alpha = max([0, min([alpha, 1])])
202
+ h1 = alpha / (alpha + 1)
203
+ h2 = (1 - alpha) / (alpha + 1)
204
+ h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
205
+ h = np.array(h)
206
+ return h
207
+
208
+
209
+ def fspecial(filter_type, *args, **kwargs):
210
+ '''
211
+ python code from:
212
+ https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
213
+ '''
214
+ if filter_type == 'gaussian':
215
+ return fspecial_gaussian(*args, **kwargs)
216
+ if filter_type == 'laplacian':
217
+ return fspecial_laplacian(*args, **kwargs)
218
+
219
+
220
+ """
221
+ # --------------------------------------------
222
+ # degradation models
223
+ # --------------------------------------------
224
+ """
225
+
226
+
227
+ def bicubic_degradation(x, sf=3):
228
+ '''
229
+ Args:
230
+ x: HxWxC image, [0, 1]
231
+ sf: down-scale factor
232
+ Return:
233
+ bicubicly downsampled LR image
234
+ '''
235
+ x = util.imresize_np(x, scale=1 / sf)
236
+ return x
237
+
238
+
239
+ def srmd_degradation(x, k, sf=3):
240
+ ''' blur + bicubic downsampling
241
+ Args:
242
+ x: HxWxC image, [0, 1]
243
+ k: hxw, double
244
+ sf: down-scale factor
245
+ Return:
246
+ downsampled LR image
247
+ Reference:
248
+ @inproceedings{zhang2018learning,
249
+ title={Learning a single convolutional super-resolution network for multiple degradations},
250
+ author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
251
+ booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
252
+ pages={3262--3271},
253
+ year={2018}
254
+ }
255
+ '''
256
+ x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror'
257
+ x = bicubic_degradation(x, sf=sf)
258
+ return x
259
+
260
+
261
+ def dpsr_degradation(x, k, sf=3):
262
+ ''' bicubic downsampling + blur
263
+ Args:
264
+ x: HxWxC image, [0, 1]
265
+ k: hxw, double
266
+ sf: down-scale factor
267
+ Return:
268
+ downsampled LR image
269
+ Reference:
270
+ @inproceedings{zhang2019deep,
271
+ title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
272
+ author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
273
+ booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
274
+ pages={1671--1681},
275
+ year={2019}
276
+ }
277
+ '''
278
+ x = bicubic_degradation(x, sf=sf)
279
+ x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
280
+ return x
281
+
282
+
283
+ def classical_degradation(x, k, sf=3):
284
+ ''' blur + downsampling
285
+ Args:
286
+ x: HxWxC image, [0, 1]/[0, 255]
287
+ k: hxw, double
288
+ sf: down-scale factor
289
+ Return:
290
+ downsampled LR image
291
+ '''
292
+ x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
293
+ # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
294
+ st = 0
295
+ return x[st::sf, st::sf, ...]
296
+
297
+
298
+ def add_sharpening(img, weight=0.5, radius=50, threshold=10):
299
+ """USM sharpening. borrowed from real-ESRGAN
300
+ Input image: I; Blurry image: B.
301
+ 1. K = I + weight * (I - B)
302
+ 2. Mask = 1 if abs(I - B) > threshold, else: 0
303
+ 3. Blur mask:
304
+ 4. Out = Mask * K + (1 - Mask) * I
305
+ Args:
306
+ img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
307
+ weight (float): Sharp weight. Default: 1.
308
+ radius (float): Kernel size of Gaussian blur. Default: 50.
309
+ threshold (int):
310
+ """
311
+ if radius % 2 == 0:
312
+ radius += 1
313
+ blur = cv2.GaussianBlur(img, (radius, radius), 0)
314
+ residual = img - blur
315
+ mask = np.abs(residual) * 255 > threshold
316
+ mask = mask.astype('float32')
317
+ soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
318
+
319
+ K = img + weight * residual
320
+ K = np.clip(K, 0, 1)
321
+ return soft_mask * K + (1 - soft_mask) * img
322
+
323
+
324
+ def add_blur(img, sf=4):
325
+ wd2 = 4.0 + sf
326
+ wd = 2.0 + 0.2 * sf
327
+
328
+ wd2 = wd2/4
329
+ wd = wd/4
330
+
331
+ if random.random() < 0.5:
332
+ l1 = wd2 * random.random()
333
+ l2 = wd2 * random.random()
334
+ k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
335
+ else:
336
+ k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random())
337
+ img = ndimage.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
338
+
339
+ return img
340
+
341
+
342
+ def add_resize(img, sf=4):
343
+ rnum = np.random.rand()
344
+ if rnum > 0.8: # up
345
+ sf1 = random.uniform(1, 2)
346
+ elif rnum < 0.7: # down
347
+ sf1 = random.uniform(0.5 / sf, 1)
348
+ else:
349
+ sf1 = 1.0
350
+ img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
351
+ img = np.clip(img, 0.0, 1.0)
352
+
353
+ return img
354
+
355
+
356
+ # def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
357
+ # noise_level = random.randint(noise_level1, noise_level2)
358
+ # rnum = np.random.rand()
359
+ # if rnum > 0.6: # add color Gaussian noise
360
+ # img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
361
+ # elif rnum < 0.4: # add grayscale Gaussian noise
362
+ # img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
363
+ # else: # add noise
364
+ # L = noise_level2 / 255.
365
+ # D = np.diag(np.random.rand(3))
366
+ # U = orth(np.random.rand(3, 3))
367
+ # conv = np.dot(np.dot(np.transpose(U), D), U)
368
+ # img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
369
+ # img = np.clip(img, 0.0, 1.0)
370
+ # return img
371
+
372
+ def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
373
+ noise_level = random.randint(noise_level1, noise_level2)
374
+ rnum = np.random.rand()
375
+ if rnum > 0.6: # add color Gaussian noise
376
+ img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
377
+ elif rnum < 0.4: # add grayscale Gaussian noise
378
+ img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
379
+ else: # add noise
380
+ L = noise_level2 / 255.
381
+ D = np.diag(np.random.rand(3))
382
+ U = orth(np.random.rand(3, 3))
383
+ conv = np.dot(np.dot(np.transpose(U), D), U)
384
+ img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
385
+ img = np.clip(img, 0.0, 1.0)
386
+ return img
387
+
388
+
389
+ def add_speckle_noise(img, noise_level1=2, noise_level2=25):
390
+ noise_level = random.randint(noise_level1, noise_level2)
391
+ img = np.clip(img, 0.0, 1.0)
392
+ rnum = random.random()
393
+ if rnum > 0.6:
394
+ img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
395
+ elif rnum < 0.4:
396
+ img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
397
+ else:
398
+ L = noise_level2 / 255.
399
+ D = np.diag(np.random.rand(3))
400
+ U = orth(np.random.rand(3, 3))
401
+ conv = np.dot(np.dot(np.transpose(U), D), U)
402
+ img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
403
+ img = np.clip(img, 0.0, 1.0)
404
+ return img
405
+
406
+
407
+ def add_Poisson_noise(img):
408
+ img = np.clip((img * 255.0).round(), 0, 255) / 255.
409
+ vals = 10 ** (2 * random.random() + 2.0) # [2, 4]
410
+ if random.random() < 0.5:
411
+ img = np.random.poisson(img * vals).astype(np.float32) / vals
412
+ else:
413
+ img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
414
+ img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
415
+ noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
416
+ img += noise_gray[:, :, np.newaxis]
417
+ img = np.clip(img, 0.0, 1.0)
418
+ return img
419
+
420
+
421
+ def add_JPEG_noise(img):
422
+ quality_factor = random.randint(80, 95)
423
+ img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
424
+ result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
425
+ img = cv2.imdecode(encimg, 1)
426
+ img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
427
+ return img
428
+
429
+
430
+ def random_crop(lq, hq, sf=4, lq_patchsize=64):
431
+ h, w = lq.shape[:2]
432
+ rnd_h = random.randint(0, h - lq_patchsize)
433
+ rnd_w = random.randint(0, w - lq_patchsize)
434
+ lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
435
+
436
+ rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
437
+ hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
438
+ return lq, hq
439
+
440
+
441
+ def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
442
+ """
443
+ This is the degradation model of BSRGAN from the paper
444
+ "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
445
+ ----------
446
+ img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
447
+ sf: scale factor
448
+ isp_model: camera ISP model
449
+ Returns
450
+ -------
451
+ img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
452
+ hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
453
+ """
454
+ isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
455
+ sf_ori = sf
456
+
457
+ h1, w1 = img.shape[:2]
458
+ img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
459
+ h, w = img.shape[:2]
460
+
461
+ if h < lq_patchsize * sf or w < lq_patchsize * sf:
462
+ raise ValueError(f'img size ({h1}X{w1}) is too small!')
463
+
464
+ hq = img.copy()
465
+
466
+ if sf == 4 and random.random() < scale2_prob: # downsample1
467
+ if np.random.rand() < 0.5:
468
+ img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
469
+ interpolation=random.choice([1, 2, 3]))
470
+ else:
471
+ img = util.imresize_np(img, 1 / 2, True)
472
+ img = np.clip(img, 0.0, 1.0)
473
+ sf = 2
474
+
475
+ shuffle_order = random.sample(range(7), 7)
476
+ idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
477
+ if idx1 > idx2: # keep downsample3 last
478
+ shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
479
+
480
+ for i in shuffle_order:
481
+
482
+ if i == 0:
483
+ img = add_blur(img, sf=sf)
484
+
485
+ elif i == 1:
486
+ img = add_blur(img, sf=sf)
487
+
488
+ elif i == 2:
489
+ a, b = img.shape[1], img.shape[0]
490
+ # downsample2
491
+ if random.random() < 0.75:
492
+ sf1 = random.uniform(1, 2 * sf)
493
+ img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
494
+ interpolation=random.choice([1, 2, 3]))
495
+ else:
496
+ k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
497
+ k_shifted = shift_pixel(k, sf)
498
+ k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
499
+ img = ndimage.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
500
+ img = img[0::sf, 0::sf, ...] # nearest downsampling
501
+ img = np.clip(img, 0.0, 1.0)
502
+
503
+ elif i == 3:
504
+ # downsample3
505
+ img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
506
+ img = np.clip(img, 0.0, 1.0)
507
+
508
+ elif i == 4:
509
+ # add Gaussian noise
510
+ img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8)
511
+
512
+ elif i == 5:
513
+ # add JPEG noise
514
+ if random.random() < jpeg_prob:
515
+ img = add_JPEG_noise(img)
516
+
517
+ elif i == 6:
518
+ # add processed camera sensor noise
519
+ if random.random() < isp_prob and isp_model is not None:
520
+ with torch.no_grad():
521
+ img, hq = isp_model.forward(img.copy(), hq)
522
+
523
+ # add final JPEG compression noise
524
+ img = add_JPEG_noise(img)
525
+
526
+ # random crop
527
+ img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
528
+
529
+ return img, hq
530
+
531
+
532
+ # todo no isp_model?
533
+ def degradation_bsrgan_variant(image, sf=4, isp_model=None, up=False):
534
+ """
535
+ This is the degradation model of BSRGAN from the paper
536
+ "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
537
+ ----------
538
+ sf: scale factor
539
+ isp_model: camera ISP model
540
+ Returns
541
+ -------
542
+ img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
543
+ hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
544
+ """
545
+ image = util.uint2single(image)
546
+ isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
547
+ sf_ori = sf
548
+
549
+ h1, w1 = image.shape[:2]
550
+ image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop
551
+ h, w = image.shape[:2]
552
+
553
+ hq = image.copy()
554
+
555
+ if sf == 4 and random.random() < scale2_prob: # downsample1
556
+ if np.random.rand() < 0.5:
557
+ image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
558
+ interpolation=random.choice([1, 2, 3]))
559
+ else:
560
+ image = util.imresize_np(image, 1 / 2, True)
561
+ image = np.clip(image, 0.0, 1.0)
562
+ sf = 2
563
+
564
+ shuffle_order = random.sample(range(7), 7)
565
+ idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
566
+ if idx1 > idx2: # keep downsample3 last
567
+ shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
568
+
569
+ for i in shuffle_order:
570
+
571
+ if i == 0:
572
+ image = add_blur(image, sf=sf)
573
+
574
+ # elif i == 1:
575
+ # image = add_blur(image, sf=sf)
576
+
577
+ if i == 0:
578
+ pass
579
+
580
+ elif i == 2:
581
+ a, b = image.shape[1], image.shape[0]
582
+ # downsample2
583
+ if random.random() < 0.8:
584
+ sf1 = random.uniform(1, 2 * sf)
585
+ image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
586
+ interpolation=random.choice([1, 2, 3]))
587
+ else:
588
+ k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
589
+ k_shifted = shift_pixel(k, sf)
590
+ k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel
591
+ image = ndimage.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
592
+ image = image[0::sf, 0::sf, ...] # nearest downsampling
593
+
594
+ image = np.clip(image, 0.0, 1.0)
595
+
596
+ elif i == 3:
597
+ # downsample3
598
+ image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
599
+ image = np.clip(image, 0.0, 1.0)
600
+
601
+ elif i == 4:
602
+ # add Gaussian noise
603
+ image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2)
604
+
605
+ elif i == 5:
606
+ # add JPEG noise
607
+ if random.random() < jpeg_prob:
608
+ image = add_JPEG_noise(image)
609
+ #
610
+ # elif i == 6:
611
+ # # add processed camera sensor noise
612
+ # if random.random() < isp_prob and isp_model is not None:
613
+ # with torch.no_grad():
614
+ # img, hq = isp_model.forward(img.copy(), hq)
615
+
616
+ # add final JPEG compression noise
617
+ image = add_JPEG_noise(image)
618
+ image = util.single2uint(image)
619
+ if up:
620
+ image = cv2.resize(image, (w1, h1), interpolation=cv2.INTER_CUBIC) # todo: random, as above? want to condition on it then
621
+ example = {"image": image}
622
+ return example
623
+
624
+
625
+
626
+
627
+ if __name__ == '__main__':
628
+ print("hey")
629
+ img = util.imread_uint('utils/test.png', 3)
630
+ img = img[:448, :448]
631
+ h = img.shape[0] // 4
632
+ print("resizing to", h)
633
+ sf = 4
634
+ deg_fn = partial(degradation_bsrgan_variant, sf=sf)
635
+ for i in range(20):
636
+ print(i)
637
+ img_hq = img
638
+ img_lq = deg_fn(img)["image"]
639
+ img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq)
640
+ print(img_lq)
641
+ img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"]
642
+ print(img_lq.shape)
643
+ print("bicubic", img_lq_bicubic.shape)
644
+ print(img_hq.shape)
645
+ lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
646
+ interpolation=0)
647
+ lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic),
648
+ (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
649
+ interpolation=0)
650
+ img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
651
+ util.imsave(img_concat, str(i) + '.png')
ldm/modules/image_degradation/utils/test.png ADDED
ldm/modules/image_degradation/utils_image.py ADDED
@@ -0,0 +1,916 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import math
3
+ import random
4
+ import numpy as np
5
+ import torch
6
+ import cv2
7
+ from torchvision.utils import make_grid
8
+ from datetime import datetime
9
+ #import matplotlib.pyplot as plt # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py
10
+
11
+
12
+ os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
13
+
14
+
15
+ '''
16
+ # --------------------------------------------
17
+ # Kai Zhang (github: https://github.com/cszn)
18
+ # 03/Mar/2019
19
+ # --------------------------------------------
20
+ # https://github.com/twhui/SRGAN-pyTorch
21
+ # https://github.com/xinntao/BasicSR
22
+ # --------------------------------------------
23
+ '''
24
+
25
+
26
+ IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif']
27
+
28
+
29
+ def is_image_file(filename):
30
+ return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
31
+
32
+
33
+ def get_timestamp():
34
+ return datetime.now().strftime('%y%m%d-%H%M%S')
35
+
36
+
37
+ def imshow(x, title=None, cbar=False, figsize=None):
38
+ plt.figure(figsize=figsize)
39
+ plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray')
40
+ if title:
41
+ plt.title(title)
42
+ if cbar:
43
+ plt.colorbar()
44
+ plt.show()
45
+
46
+
47
+ def surf(Z, cmap='rainbow', figsize=None):
48
+ plt.figure(figsize=figsize)
49
+ ax3 = plt.axes(projection='3d')
50
+
51
+ w, h = Z.shape[:2]
52
+ xx = np.arange(0,w,1)
53
+ yy = np.arange(0,h,1)
54
+ X, Y = np.meshgrid(xx, yy)
55
+ ax3.plot_surface(X,Y,Z,cmap=cmap)
56
+ #ax3.contour(X,Y,Z, zdim='z',offset=-2,cmap=cmap)
57
+ plt.show()
58
+
59
+
60
+ '''
61
+ # --------------------------------------------
62
+ # get image pathes
63
+ # --------------------------------------------
64
+ '''
65
+
66
+
67
+ def get_image_paths(dataroot):
68
+ paths = None # return None if dataroot is None
69
+ if dataroot is not None:
70
+ paths = sorted(_get_paths_from_images(dataroot))
71
+ return paths
72
+
73
+
74
+ def _get_paths_from_images(path):
75
+ assert os.path.isdir(path), '{:s} is not a valid directory'.format(path)
76
+ images = []
77
+ for dirpath, _, fnames in sorted(os.walk(path)):
78
+ for fname in sorted(fnames):
79
+ if is_image_file(fname):
80
+ img_path = os.path.join(dirpath, fname)
81
+ images.append(img_path)
82
+ assert images, '{:s} has no valid image file'.format(path)
83
+ return images
84
+
85
+
86
+ '''
87
+ # --------------------------------------------
88
+ # split large images into small images
89
+ # --------------------------------------------
90
+ '''
91
+
92
+
93
+ def patches_from_image(img, p_size=512, p_overlap=64, p_max=800):
94
+ w, h = img.shape[:2]
95
+ patches = []
96
+ if w > p_max and h > p_max:
97
+ w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int))
98
+ h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int))
99
+ w1.append(w-p_size)
100
+ h1.append(h-p_size)
101
+ # print(w1)
102
+ # print(h1)
103
+ for i in w1:
104
+ for j in h1:
105
+ patches.append(img[i:i+p_size, j:j+p_size,:])
106
+ else:
107
+ patches.append(img)
108
+
109
+ return patches
110
+
111
+
112
+ def imssave(imgs, img_path):
113
+ """
114
+ imgs: list, N images of size WxHxC
115
+ """
116
+ img_name, ext = os.path.splitext(os.path.basename(img_path))
117
+
118
+ for i, img in enumerate(imgs):
119
+ if img.ndim == 3:
120
+ img = img[:, :, [2, 1, 0]]
121
+ new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png')
122
+ cv2.imwrite(new_path, img)
123
+
124
+
125
+ def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000):
126
+ """
127
+ split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size),
128
+ and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max)
129
+ will be splitted.
130
+ Args:
131
+ original_dataroot:
132
+ taget_dataroot:
133
+ p_size: size of small images
134
+ p_overlap: patch size in training is a good choice
135
+ p_max: images with smaller size than (p_max)x(p_max) keep unchanged.
136
+ """
137
+ paths = get_image_paths(original_dataroot)
138
+ for img_path in paths:
139
+ # img_name, ext = os.path.splitext(os.path.basename(img_path))
140
+ img = imread_uint(img_path, n_channels=n_channels)
141
+ patches = patches_from_image(img, p_size, p_overlap, p_max)
142
+ imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path)))
143
+ #if original_dataroot == taget_dataroot:
144
+ #del img_path
145
+
146
+ '''
147
+ # --------------------------------------------
148
+ # makedir
149
+ # --------------------------------------------
150
+ '''
151
+
152
+
153
+ def mkdir(path):
154
+ if not os.path.exists(path):
155
+ os.makedirs(path)
156
+
157
+
158
+ def mkdirs(paths):
159
+ if isinstance(paths, str):
160
+ mkdir(paths)
161
+ else:
162
+ for path in paths:
163
+ mkdir(path)
164
+
165
+
166
+ def mkdir_and_rename(path):
167
+ if os.path.exists(path):
168
+ new_name = path + '_archived_' + get_timestamp()
169
+ print('Path already exists. Rename it to [{:s}]'.format(new_name))
170
+ os.rename(path, new_name)
171
+ os.makedirs(path)
172
+
173
+
174
+ '''
175
+ # --------------------------------------------
176
+ # read image from path
177
+ # opencv is fast, but read BGR numpy image
178
+ # --------------------------------------------
179
+ '''
180
+
181
+
182
+ # --------------------------------------------
183
+ # get uint8 image of size HxWxn_channles (RGB)
184
+ # --------------------------------------------
185
+ def imread_uint(path, n_channels=3):
186
+ # input: path
187
+ # output: HxWx3(RGB or GGG), or HxWx1 (G)
188
+ if n_channels == 1:
189
+ img = cv2.imread(path, 0) # cv2.IMREAD_GRAYSCALE
190
+ img = np.expand_dims(img, axis=2) # HxWx1
191
+ elif n_channels == 3:
192
+ img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # BGR or G
193
+ if img.ndim == 2:
194
+ img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) # GGG
195
+ else:
196
+ img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # RGB
197
+ return img
198
+
199
+
200
+ # --------------------------------------------
201
+ # matlab's imwrite
202
+ # --------------------------------------------
203
+ def imsave(img, img_path):
204
+ img = np.squeeze(img)
205
+ if img.ndim == 3:
206
+ img = img[:, :, [2, 1, 0]]
207
+ cv2.imwrite(img_path, img)
208
+
209
+ def imwrite(img, img_path):
210
+ img = np.squeeze(img)
211
+ if img.ndim == 3:
212
+ img = img[:, :, [2, 1, 0]]
213
+ cv2.imwrite(img_path, img)
214
+
215
+
216
+
217
+ # --------------------------------------------
218
+ # get single image of size HxWxn_channles (BGR)
219
+ # --------------------------------------------
220
+ def read_img(path):
221
+ # read image by cv2
222
+ # return: Numpy float32, HWC, BGR, [0,1]
223
+ img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # cv2.IMREAD_GRAYSCALE
224
+ img = img.astype(np.float32) / 255.
225
+ if img.ndim == 2:
226
+ img = np.expand_dims(img, axis=2)
227
+ # some images have 4 channels
228
+ if img.shape[2] > 3:
229
+ img = img[:, :, :3]
230
+ return img
231
+
232
+
233
+ '''
234
+ # --------------------------------------------
235
+ # image format conversion
236
+ # --------------------------------------------
237
+ # numpy(single) <---> numpy(unit)
238
+ # numpy(single) <---> tensor
239
+ # numpy(unit) <---> tensor
240
+ # --------------------------------------------
241
+ '''
242
+
243
+
244
+ # --------------------------------------------
245
+ # numpy(single) [0, 1] <---> numpy(unit)
246
+ # --------------------------------------------
247
+
248
+
249
+ def uint2single(img):
250
+
251
+ return np.float32(img/255.)
252
+
253
+
254
+ def single2uint(img):
255
+
256
+ return np.uint8((img.clip(0, 1)*255.).round())
257
+
258
+
259
+ def uint162single(img):
260
+
261
+ return np.float32(img/65535.)
262
+
263
+
264
+ def single2uint16(img):
265
+
266
+ return np.uint16((img.clip(0, 1)*65535.).round())
267
+
268
+
269
+ # --------------------------------------------
270
+ # numpy(unit) (HxWxC or HxW) <---> tensor
271
+ # --------------------------------------------
272
+
273
+
274
+ # convert uint to 4-dimensional torch tensor
275
+ def uint2tensor4(img):
276
+ if img.ndim == 2:
277
+ img = np.expand_dims(img, axis=2)
278
+ return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0)
279
+
280
+
281
+ # convert uint to 3-dimensional torch tensor
282
+ def uint2tensor3(img):
283
+ if img.ndim == 2:
284
+ img = np.expand_dims(img, axis=2)
285
+ return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.)
286
+
287
+
288
+ # convert 2/3/4-dimensional torch tensor to uint
289
+ def tensor2uint(img):
290
+ img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy()
291
+ if img.ndim == 3:
292
+ img = np.transpose(img, (1, 2, 0))
293
+ return np.uint8((img*255.0).round())
294
+
295
+
296
+ # --------------------------------------------
297
+ # numpy(single) (HxWxC) <---> tensor
298
+ # --------------------------------------------
299
+
300
+
301
+ # convert single (HxWxC) to 3-dimensional torch tensor
302
+ def single2tensor3(img):
303
+ return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float()
304
+
305
+
306
+ # convert single (HxWxC) to 4-dimensional torch tensor
307
+ def single2tensor4(img):
308
+ return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0)
309
+
310
+
311
+ # convert torch tensor to single
312
+ def tensor2single(img):
313
+ img = img.data.squeeze().float().cpu().numpy()
314
+ if img.ndim == 3:
315
+ img = np.transpose(img, (1, 2, 0))
316
+
317
+ return img
318
+
319
+ # convert torch tensor to single
320
+ def tensor2single3(img):
321
+ img = img.data.squeeze().float().cpu().numpy()
322
+ if img.ndim == 3:
323
+ img = np.transpose(img, (1, 2, 0))
324
+ elif img.ndim == 2:
325
+ img = np.expand_dims(img, axis=2)
326
+ return img
327
+
328
+
329
+ def single2tensor5(img):
330
+ return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0)
331
+
332
+
333
+ def single32tensor5(img):
334
+ return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0)
335
+
336
+
337
+ def single42tensor4(img):
338
+ return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float()
339
+
340
+
341
+ # from skimage.io import imread, imsave
342
+ def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)):
343
+ '''
344
+ Converts a torch Tensor into an image Numpy array of BGR channel order
345
+ Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order
346
+ Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default)
347
+ '''
348
+ tensor = tensor.squeeze().float().cpu().clamp_(*min_max) # squeeze first, then clamp
349
+ tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0]) # to range [0,1]
350
+ n_dim = tensor.dim()
351
+ if n_dim == 4:
352
+ n_img = len(tensor)
353
+ img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy()
354
+ img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR
355
+ elif n_dim == 3:
356
+ img_np = tensor.numpy()
357
+ img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR
358
+ elif n_dim == 2:
359
+ img_np = tensor.numpy()
360
+ else:
361
+ raise TypeError(
362
+ 'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim))
363
+ if out_type == np.uint8:
364
+ img_np = (img_np * 255.0).round()
365
+ # Important. Unlike matlab, numpy.unit8() WILL NOT round by default.
366
+ return img_np.astype(out_type)
367
+
368
+
369
+ '''
370
+ # --------------------------------------------
371
+ # Augmentation, flipe and/or rotate
372
+ # --------------------------------------------
373
+ # The following two are enough.
374
+ # (1) augmet_img: numpy image of WxHxC or WxH
375
+ # (2) augment_img_tensor4: tensor image 1xCxWxH
376
+ # --------------------------------------------
377
+ '''
378
+
379
+
380
+ def augment_img(img, mode=0):
381
+ '''Kai Zhang (github: https://github.com/cszn)
382
+ '''
383
+ if mode == 0:
384
+ return img
385
+ elif mode == 1:
386
+ return np.flipud(np.rot90(img))
387
+ elif mode == 2:
388
+ return np.flipud(img)
389
+ elif mode == 3:
390
+ return np.rot90(img, k=3)
391
+ elif mode == 4:
392
+ return np.flipud(np.rot90(img, k=2))
393
+ elif mode == 5:
394
+ return np.rot90(img)
395
+ elif mode == 6:
396
+ return np.rot90(img, k=2)
397
+ elif mode == 7:
398
+ return np.flipud(np.rot90(img, k=3))
399
+
400
+
401
+ def augment_img_tensor4(img, mode=0):
402
+ '''Kai Zhang (github: https://github.com/cszn)
403
+ '''
404
+ if mode == 0:
405
+ return img
406
+ elif mode == 1:
407
+ return img.rot90(1, [2, 3]).flip([2])
408
+ elif mode == 2:
409
+ return img.flip([2])
410
+ elif mode == 3:
411
+ return img.rot90(3, [2, 3])
412
+ elif mode == 4:
413
+ return img.rot90(2, [2, 3]).flip([2])
414
+ elif mode == 5:
415
+ return img.rot90(1, [2, 3])
416
+ elif mode == 6:
417
+ return img.rot90(2, [2, 3])
418
+ elif mode == 7:
419
+ return img.rot90(3, [2, 3]).flip([2])
420
+
421
+
422
+ def augment_img_tensor(img, mode=0):
423
+ '''Kai Zhang (github: https://github.com/cszn)
424
+ '''
425
+ img_size = img.size()
426
+ img_np = img.data.cpu().numpy()
427
+ if len(img_size) == 3:
428
+ img_np = np.transpose(img_np, (1, 2, 0))
429
+ elif len(img_size) == 4:
430
+ img_np = np.transpose(img_np, (2, 3, 1, 0))
431
+ img_np = augment_img(img_np, mode=mode)
432
+ img_tensor = torch.from_numpy(np.ascontiguousarray(img_np))
433
+ if len(img_size) == 3:
434
+ img_tensor = img_tensor.permute(2, 0, 1)
435
+ elif len(img_size) == 4:
436
+ img_tensor = img_tensor.permute(3, 2, 0, 1)
437
+
438
+ return img_tensor.type_as(img)
439
+
440
+
441
+ def augment_img_np3(img, mode=0):
442
+ if mode == 0:
443
+ return img
444
+ elif mode == 1:
445
+ return img.transpose(1, 0, 2)
446
+ elif mode == 2:
447
+ return img[::-1, :, :]
448
+ elif mode == 3:
449
+ img = img[::-1, :, :]
450
+ img = img.transpose(1, 0, 2)
451
+ return img
452
+ elif mode == 4:
453
+ return img[:, ::-1, :]
454
+ elif mode == 5:
455
+ img = img[:, ::-1, :]
456
+ img = img.transpose(1, 0, 2)
457
+ return img
458
+ elif mode == 6:
459
+ img = img[:, ::-1, :]
460
+ img = img[::-1, :, :]
461
+ return img
462
+ elif mode == 7:
463
+ img = img[:, ::-1, :]
464
+ img = img[::-1, :, :]
465
+ img = img.transpose(1, 0, 2)
466
+ return img
467
+
468
+
469
+ def augment_imgs(img_list, hflip=True, rot=True):
470
+ # horizontal flip OR rotate
471
+ hflip = hflip and random.random() < 0.5
472
+ vflip = rot and random.random() < 0.5
473
+ rot90 = rot and random.random() < 0.5
474
+
475
+ def _augment(img):
476
+ if hflip:
477
+ img = img[:, ::-1, :]
478
+ if vflip:
479
+ img = img[::-1, :, :]
480
+ if rot90:
481
+ img = img.transpose(1, 0, 2)
482
+ return img
483
+
484
+ return [_augment(img) for img in img_list]
485
+
486
+
487
+ '''
488
+ # --------------------------------------------
489
+ # modcrop and shave
490
+ # --------------------------------------------
491
+ '''
492
+
493
+
494
+ def modcrop(img_in, scale):
495
+ # img_in: Numpy, HWC or HW
496
+ img = np.copy(img_in)
497
+ if img.ndim == 2:
498
+ H, W = img.shape
499
+ H_r, W_r = H % scale, W % scale
500
+ img = img[:H - H_r, :W - W_r]
501
+ elif img.ndim == 3:
502
+ H, W, C = img.shape
503
+ H_r, W_r = H % scale, W % scale
504
+ img = img[:H - H_r, :W - W_r, :]
505
+ else:
506
+ raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim))
507
+ return img
508
+
509
+
510
+ def shave(img_in, border=0):
511
+ # img_in: Numpy, HWC or HW
512
+ img = np.copy(img_in)
513
+ h, w = img.shape[:2]
514
+ img = img[border:h-border, border:w-border]
515
+ return img
516
+
517
+
518
+ '''
519
+ # --------------------------------------------
520
+ # image processing process on numpy image
521
+ # channel_convert(in_c, tar_type, img_list):
522
+ # rgb2ycbcr(img, only_y=True):
523
+ # bgr2ycbcr(img, only_y=True):
524
+ # ycbcr2rgb(img):
525
+ # --------------------------------------------
526
+ '''
527
+
528
+
529
+ def rgb2ycbcr(img, only_y=True):
530
+ '''same as matlab rgb2ycbcr
531
+ only_y: only return Y channel
532
+ Input:
533
+ uint8, [0, 255]
534
+ float, [0, 1]
535
+ '''
536
+ in_img_type = img.dtype
537
+ img.astype(np.float32)
538
+ if in_img_type != np.uint8:
539
+ img *= 255.
540
+ # convert
541
+ if only_y:
542
+ rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0
543
+ else:
544
+ rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786],
545
+ [24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128]
546
+ if in_img_type == np.uint8:
547
+ rlt = rlt.round()
548
+ else:
549
+ rlt /= 255.
550
+ return rlt.astype(in_img_type)
551
+
552
+
553
+ def ycbcr2rgb(img):
554
+ '''same as matlab ycbcr2rgb
555
+ Input:
556
+ uint8, [0, 255]
557
+ float, [0, 1]
558
+ '''
559
+ in_img_type = img.dtype
560
+ img.astype(np.float32)
561
+ if in_img_type != np.uint8:
562
+ img *= 255.
563
+ # convert
564
+ rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071],
565
+ [0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836]
566
+ if in_img_type == np.uint8:
567
+ rlt = rlt.round()
568
+ else:
569
+ rlt /= 255.
570
+ return rlt.astype(in_img_type)
571
+
572
+
573
+ def bgr2ycbcr(img, only_y=True):
574
+ '''bgr version of rgb2ycbcr
575
+ only_y: only return Y channel
576
+ Input:
577
+ uint8, [0, 255]
578
+ float, [0, 1]
579
+ '''
580
+ in_img_type = img.dtype
581
+ img.astype(np.float32)
582
+ if in_img_type != np.uint8:
583
+ img *= 255.
584
+ # convert
585
+ if only_y:
586
+ rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0
587
+ else:
588
+ rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786],
589
+ [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128]
590
+ if in_img_type == np.uint8:
591
+ rlt = rlt.round()
592
+ else:
593
+ rlt /= 255.
594
+ return rlt.astype(in_img_type)
595
+
596
+
597
+ def channel_convert(in_c, tar_type, img_list):
598
+ # conversion among BGR, gray and y
599
+ if in_c == 3 and tar_type == 'gray': # BGR to gray
600
+ gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list]
601
+ return [np.expand_dims(img, axis=2) for img in gray_list]
602
+ elif in_c == 3 and tar_type == 'y': # BGR to y
603
+ y_list = [bgr2ycbcr(img, only_y=True) for img in img_list]
604
+ return [np.expand_dims(img, axis=2) for img in y_list]
605
+ elif in_c == 1 and tar_type == 'RGB': # gray/y to BGR
606
+ return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list]
607
+ else:
608
+ return img_list
609
+
610
+
611
+ '''
612
+ # --------------------------------------------
613
+ # metric, PSNR and SSIM
614
+ # --------------------------------------------
615
+ '''
616
+
617
+
618
+ # --------------------------------------------
619
+ # PSNR
620
+ # --------------------------------------------
621
+ def calculate_psnr(img1, img2, border=0):
622
+ # img1 and img2 have range [0, 255]
623
+ #img1 = img1.squeeze()
624
+ #img2 = img2.squeeze()
625
+ if not img1.shape == img2.shape:
626
+ raise ValueError('Input images must have the same dimensions.')
627
+ h, w = img1.shape[:2]
628
+ img1 = img1[border:h-border, border:w-border]
629
+ img2 = img2[border:h-border, border:w-border]
630
+
631
+ img1 = img1.astype(np.float64)
632
+ img2 = img2.astype(np.float64)
633
+ mse = np.mean((img1 - img2)**2)
634
+ if mse == 0:
635
+ return float('inf')
636
+ return 20 * math.log10(255.0 / math.sqrt(mse))
637
+
638
+
639
+ # --------------------------------------------
640
+ # SSIM
641
+ # --------------------------------------------
642
+ def calculate_ssim(img1, img2, border=0):
643
+ '''calculate SSIM
644
+ the same outputs as MATLAB's
645
+ img1, img2: [0, 255]
646
+ '''
647
+ #img1 = img1.squeeze()
648
+ #img2 = img2.squeeze()
649
+ if not img1.shape == img2.shape:
650
+ raise ValueError('Input images must have the same dimensions.')
651
+ h, w = img1.shape[:2]
652
+ img1 = img1[border:h-border, border:w-border]
653
+ img2 = img2[border:h-border, border:w-border]
654
+
655
+ if img1.ndim == 2:
656
+ return ssim(img1, img2)
657
+ elif img1.ndim == 3:
658
+ if img1.shape[2] == 3:
659
+ ssims = []
660
+ for i in range(3):
661
+ ssims.append(ssim(img1[:,:,i], img2[:,:,i]))
662
+ return np.array(ssims).mean()
663
+ elif img1.shape[2] == 1:
664
+ return ssim(np.squeeze(img1), np.squeeze(img2))
665
+ else:
666
+ raise ValueError('Wrong input image dimensions.')
667
+
668
+
669
+ def ssim(img1, img2):
670
+ C1 = (0.01 * 255)**2
671
+ C2 = (0.03 * 255)**2
672
+
673
+ img1 = img1.astype(np.float64)
674
+ img2 = img2.astype(np.float64)
675
+ kernel = cv2.getGaussianKernel(11, 1.5)
676
+ window = np.outer(kernel, kernel.transpose())
677
+
678
+ mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid
679
+ mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
680
+ mu1_sq = mu1**2
681
+ mu2_sq = mu2**2
682
+ mu1_mu2 = mu1 * mu2
683
+ sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq
684
+ sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq
685
+ sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
686
+
687
+ ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *
688
+ (sigma1_sq + sigma2_sq + C2))
689
+ return ssim_map.mean()
690
+
691
+
692
+ '''
693
+ # --------------------------------------------
694
+ # matlab's bicubic imresize (numpy and torch) [0, 1]
695
+ # --------------------------------------------
696
+ '''
697
+
698
+
699
+ # matlab 'imresize' function, now only support 'bicubic'
700
+ def cubic(x):
701
+ absx = torch.abs(x)
702
+ absx2 = absx**2
703
+ absx3 = absx**3
704
+ return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \
705
+ (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx))
706
+
707
+
708
+ def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing):
709
+ if (scale < 1) and (antialiasing):
710
+ # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width
711
+ kernel_width = kernel_width / scale
712
+
713
+ # Output-space coordinates
714
+ x = torch.linspace(1, out_length, out_length)
715
+
716
+ # Input-space coordinates. Calculate the inverse mapping such that 0.5
717
+ # in output space maps to 0.5 in input space, and 0.5+scale in output
718
+ # space maps to 1.5 in input space.
719
+ u = x / scale + 0.5 * (1 - 1 / scale)
720
+
721
+ # What is the left-most pixel that can be involved in the computation?
722
+ left = torch.floor(u - kernel_width / 2)
723
+
724
+ # What is the maximum number of pixels that can be involved in the
725
+ # computation? Note: it's OK to use an extra pixel here; if the
726
+ # corresponding weights are all zero, it will be eliminated at the end
727
+ # of this function.
728
+ P = math.ceil(kernel_width) + 2
729
+
730
+ # The indices of the input pixels involved in computing the k-th output
731
+ # pixel are in row k of the indices matrix.
732
+ indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view(
733
+ 1, P).expand(out_length, P)
734
+
735
+ # The weights used to compute the k-th output pixel are in row k of the
736
+ # weights matrix.
737
+ distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices
738
+ # apply cubic kernel
739
+ if (scale < 1) and (antialiasing):
740
+ weights = scale * cubic(distance_to_center * scale)
741
+ else:
742
+ weights = cubic(distance_to_center)
743
+ # Normalize the weights matrix so that each row sums to 1.
744
+ weights_sum = torch.sum(weights, 1).view(out_length, 1)
745
+ weights = weights / weights_sum.expand(out_length, P)
746
+
747
+ # If a column in weights is all zero, get rid of it. only consider the first and last column.
748
+ weights_zero_tmp = torch.sum((weights == 0), 0)
749
+ if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6):
750
+ indices = indices.narrow(1, 1, P - 2)
751
+ weights = weights.narrow(1, 1, P - 2)
752
+ if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6):
753
+ indices = indices.narrow(1, 0, P - 2)
754
+ weights = weights.narrow(1, 0, P - 2)
755
+ weights = weights.contiguous()
756
+ indices = indices.contiguous()
757
+ sym_len_s = -indices.min() + 1
758
+ sym_len_e = indices.max() - in_length
759
+ indices = indices + sym_len_s - 1
760
+ return weights, indices, int(sym_len_s), int(sym_len_e)
761
+
762
+
763
+ # --------------------------------------------
764
+ # imresize for tensor image [0, 1]
765
+ # --------------------------------------------
766
+ def imresize(img, scale, antialiasing=True):
767
+ # Now the scale should be the same for H and W
768
+ # input: img: pytorch tensor, CHW or HW [0,1]
769
+ # output: CHW or HW [0,1] w/o round
770
+ need_squeeze = True if img.dim() == 2 else False
771
+ if need_squeeze:
772
+ img.unsqueeze_(0)
773
+ in_C, in_H, in_W = img.size()
774
+ out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
775
+ kernel_width = 4
776
+ kernel = 'cubic'
777
+
778
+ # Return the desired dimension order for performing the resize. The
779
+ # strategy is to perform the resize first along the dimension with the
780
+ # smallest scale factor.
781
+ # Now we do not support this.
782
+
783
+ # get weights and indices
784
+ weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
785
+ in_H, out_H, scale, kernel, kernel_width, antialiasing)
786
+ weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
787
+ in_W, out_W, scale, kernel, kernel_width, antialiasing)
788
+ # process H dimension
789
+ # symmetric copying
790
+ img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W)
791
+ img_aug.narrow(1, sym_len_Hs, in_H).copy_(img)
792
+
793
+ sym_patch = img[:, :sym_len_Hs, :]
794
+ inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
795
+ sym_patch_inv = sym_patch.index_select(1, inv_idx)
796
+ img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv)
797
+
798
+ sym_patch = img[:, -sym_len_He:, :]
799
+ inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
800
+ sym_patch_inv = sym_patch.index_select(1, inv_idx)
801
+ img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
802
+
803
+ out_1 = torch.FloatTensor(in_C, out_H, in_W)
804
+ kernel_width = weights_H.size(1)
805
+ for i in range(out_H):
806
+ idx = int(indices_H[i][0])
807
+ for j in range(out_C):
808
+ out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i])
809
+
810
+ # process W dimension
811
+ # symmetric copying
812
+ out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We)
813
+ out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1)
814
+
815
+ sym_patch = out_1[:, :, :sym_len_Ws]
816
+ inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
817
+ sym_patch_inv = sym_patch.index_select(2, inv_idx)
818
+ out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv)
819
+
820
+ sym_patch = out_1[:, :, -sym_len_We:]
821
+ inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
822
+ sym_patch_inv = sym_patch.index_select(2, inv_idx)
823
+ out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
824
+
825
+ out_2 = torch.FloatTensor(in_C, out_H, out_W)
826
+ kernel_width = weights_W.size(1)
827
+ for i in range(out_W):
828
+ idx = int(indices_W[i][0])
829
+ for j in range(out_C):
830
+ out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i])
831
+ if need_squeeze:
832
+ out_2.squeeze_()
833
+ return out_2
834
+
835
+
836
+ # --------------------------------------------
837
+ # imresize for numpy image [0, 1]
838
+ # --------------------------------------------
839
+ def imresize_np(img, scale, antialiasing=True):
840
+ # Now the scale should be the same for H and W
841
+ # input: img: Numpy, HWC or HW [0,1]
842
+ # output: HWC or HW [0,1] w/o round
843
+ img = torch.from_numpy(img)
844
+ need_squeeze = True if img.dim() == 2 else False
845
+ if need_squeeze:
846
+ img.unsqueeze_(2)
847
+
848
+ in_H, in_W, in_C = img.size()
849
+ out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
850
+ kernel_width = 4
851
+ kernel = 'cubic'
852
+
853
+ # Return the desired dimension order for performing the resize. The
854
+ # strategy is to perform the resize first along the dimension with the
855
+ # smallest scale factor.
856
+ # Now we do not support this.
857
+
858
+ # get weights and indices
859
+ weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
860
+ in_H, out_H, scale, kernel, kernel_width, antialiasing)
861
+ weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
862
+ in_W, out_W, scale, kernel, kernel_width, antialiasing)
863
+ # process H dimension
864
+ # symmetric copying
865
+ img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C)
866
+ img_aug.narrow(0, sym_len_Hs, in_H).copy_(img)
867
+
868
+ sym_patch = img[:sym_len_Hs, :, :]
869
+ inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
870
+ sym_patch_inv = sym_patch.index_select(0, inv_idx)
871
+ img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv)
872
+
873
+ sym_patch = img[-sym_len_He:, :, :]
874
+ inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
875
+ sym_patch_inv = sym_patch.index_select(0, inv_idx)
876
+ img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
877
+
878
+ out_1 = torch.FloatTensor(out_H, in_W, in_C)
879
+ kernel_width = weights_H.size(1)
880
+ for i in range(out_H):
881
+ idx = int(indices_H[i][0])
882
+ for j in range(out_C):
883
+ out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i])
884
+
885
+ # process W dimension
886
+ # symmetric copying
887
+ out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C)
888
+ out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1)
889
+
890
+ sym_patch = out_1[:, :sym_len_Ws, :]
891
+ inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
892
+ sym_patch_inv = sym_patch.index_select(1, inv_idx)
893
+ out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv)
894
+
895
+ sym_patch = out_1[:, -sym_len_We:, :]
896
+ inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
897
+ sym_patch_inv = sym_patch.index_select(1, inv_idx)
898
+ out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
899
+
900
+ out_2 = torch.FloatTensor(out_H, out_W, in_C)
901
+ kernel_width = weights_W.size(1)
902
+ for i in range(out_W):
903
+ idx = int(indices_W[i][0])
904
+ for j in range(out_C):
905
+ out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i])
906
+ if need_squeeze:
907
+ out_2.squeeze_()
908
+
909
+ return out_2.numpy()
910
+
911
+
912
+ if __name__ == '__main__':
913
+ print('---')
914
+ # img = imread_uint('test.bmp', 3)
915
+ # img = uint2single(img)
916
+ # img_bicubic = imresize_np(img, 1/4)
ldm/modules/image_encoders/__init__.py ADDED
File without changes
ldm/modules/image_encoders/modules.py ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch.nn as nn
2
+ from transformers import CLIPVisionModel
3
+ from .xf import LayerNorm, Transformer
4
+
5
+ class AbstractEncoder(nn.Module):
6
+ def __init__(self):
7
+ super().__init__()
8
+
9
+ def encode(self, *args, **kwargs):
10
+ raise NotImplementedError
11
+
12
+ class FrozenCLIPImageEmbedder(AbstractEncoder):
13
+ """Uses the CLIP transformer encoder for text (from Hugging Face)"""
14
+ def __init__(self, version="openai/clip-vit-large-patch14"):
15
+ super().__init__()
16
+ self.transformer = CLIPVisionModel.from_pretrained(version)
17
+ self.final_ln = LayerNorm(1024)
18
+ self.mapper = Transformer(
19
+ 1,
20
+ 1024,
21
+ 5,
22
+ 1,
23
+ )
24
+
25
+ self.freeze()
26
+
27
+ def freeze(self):
28
+ self.transformer = self.transformer.eval()
29
+ for param in self.parameters():
30
+ param.requires_grad = False
31
+ for param in self.mapper.parameters():
32
+ param.requires_grad = True
33
+ for param in self.final_ln.parameters():
34
+ param.requires_grad = True
35
+
36
+ def forward(self, image):
37
+ outputs = self.transformer(pixel_values=image)
38
+ z = outputs.pooler_output
39
+ z = z.unsqueeze(1)
40
+ z = self.mapper(z)
41
+ z = self.final_ln(z)
42
+ return z
43
+
44
+ def encode(self, image):
45
+ if isinstance(image, list):
46
+ image = image[0]
47
+ return self(image)
48
+
49
+ if __name__ == "__main__":
50
+ from ldm.util import count_params
51
+ model = FrozenCLIPImageEmbedder()
52
+ count_params(model, verbose=True)
ldm/modules/image_encoders/xf.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Transformer implementation adapted from CLIP ViT:
3
+ https://github.com/openai/CLIP/blob/4c0275784d6d9da97ca1f47eaaee31de1867da91/clip/model.py
4
+ """
5
+
6
+ import math
7
+
8
+ import torch as th
9
+ import torch.nn as nn
10
+
11
+
12
+ def convert_module_to_f16(l):
13
+ """
14
+ Convert primitive modules to float16.
15
+ """
16
+ if isinstance(l, (nn.Linear, nn.Conv2d, nn.ConvTranspose2d)):
17
+ l.weight.data = l.weight.data.half()
18
+ if l.bias is not None:
19
+ l.bias.data = l.bias.data.half()
20
+
21
+
22
+ class LayerNorm(nn.LayerNorm):
23
+ """
24
+ Implementation that supports fp16 inputs but fp32 gains/biases.
25
+ """
26
+
27
+ def forward(self, x: th.Tensor):
28
+ return super().forward(x.float()).to(x.dtype)
29
+
30
+
31
+ class MultiheadAttention(nn.Module):
32
+ def __init__(self, n_ctx, width, heads):
33
+ super().__init__()
34
+ self.n_ctx = n_ctx
35
+ self.width = width
36
+ self.heads = heads
37
+ self.c_qkv = nn.Linear(width, width * 3)
38
+ self.c_proj = nn.Linear(width, width)
39
+ self.attention = QKVMultiheadAttention(heads, n_ctx)
40
+
41
+ def forward(self, x):
42
+ x = self.c_qkv(x)
43
+ x = self.attention(x)
44
+ x = self.c_proj(x)
45
+ return x
46
+
47
+
48
+ class MLP(nn.Module):
49
+ def __init__(self, width):
50
+ super().__init__()
51
+ self.width = width
52
+ self.c_fc = nn.Linear(width, width * 4)
53
+ self.c_proj = nn.Linear(width * 4, width)
54
+ self.gelu = nn.GELU()
55
+
56
+ def forward(self, x):
57
+ return self.c_proj(self.gelu(self.c_fc(x)))
58
+
59
+
60
+ class QKVMultiheadAttention(nn.Module):
61
+ def __init__(self, n_heads: int, n_ctx: int):
62
+ super().__init__()
63
+ self.n_heads = n_heads
64
+ self.n_ctx = n_ctx
65
+
66
+ def forward(self, qkv):
67
+ bs, n_ctx, width = qkv.shape
68
+ attn_ch = width // self.n_heads // 3
69
+ scale = 1 / math.sqrt(math.sqrt(attn_ch))
70
+ qkv = qkv.view(bs, n_ctx, self.n_heads, -1)
71
+ q, k, v = th.split(qkv, attn_ch, dim=-1)
72
+ weight = th.einsum(
73
+ "bthc,bshc->bhts", q * scale, k * scale
74
+ ) # More stable with f16 than dividing afterwards
75
+ wdtype = weight.dtype
76
+ weight = th.softmax(weight.float(), dim=-1).type(wdtype)
77
+ return th.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
78
+
79
+
80
+ class ResidualAttentionBlock(nn.Module):
81
+ def __init__(
82
+ self,
83
+ n_ctx: int,
84
+ width: int,
85
+ heads: int,
86
+ ):
87
+ super().__init__()
88
+
89
+ self.attn = MultiheadAttention(
90
+ n_ctx,
91
+ width,
92
+ heads,
93
+ )
94
+ self.ln_1 = LayerNorm(width)
95
+ self.mlp = MLP(width)
96
+ self.ln_2 = LayerNorm(width)
97
+
98
+ def forward(self, x: th.Tensor):
99
+ x = x + self.attn(self.ln_1(x))
100
+ x = x + self.mlp(self.ln_2(x))
101
+ return x
102
+
103
+
104
+ class Transformer(nn.Module):
105
+ def __init__(
106
+ self,
107
+ n_ctx: int,
108
+ width: int,
109
+ layers: int,
110
+ heads: int,
111
+ ):
112
+ super().__init__()
113
+ self.n_ctx = n_ctx
114
+ self.width = width
115
+ self.layers = layers
116
+ self.resblocks = nn.ModuleList(
117
+ [
118
+ ResidualAttentionBlock(
119
+ n_ctx,
120
+ width,
121
+ heads,
122
+ )
123
+ for _ in range(layers)
124
+ ]
125
+ )
126
+
127
+ def forward(self, x: th.Tensor):
128
+ for block in self.resblocks:
129
+ x = block(x)
130
+ return x
ldm/modules/midas/__init__.py ADDED
File without changes
ldm/modules/midas/api.py ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # based on https://github.com/isl-org/MiDaS
2
+
3
+ import cv2
4
+ import torch
5
+ import torch.nn as nn
6
+ from torchvision.transforms import Compose
7
+
8
+ from ldm.modules.midas.midas.dpt_depth import DPTDepthModel
9
+ from ldm.modules.midas.midas.midas_net import MidasNet
10
+ from ldm.modules.midas.midas.midas_net_custom import MidasNet_small
11
+ from ldm.modules.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet
12
+
13
+
14
+ ISL_PATHS = {
15
+ "dpt_large": "midas_models/dpt_large-midas-2f21e586.pt",
16
+ "dpt_hybrid": "midas_models/dpt_hybrid-midas-501f0c75.pt",
17
+ "midas_v21": "",
18
+ "midas_v21_small": "",
19
+ }
20
+
21
+
22
+ def disabled_train(self, mode=True):
23
+ """Overwrite model.train with this function to make sure train/eval mode
24
+ does not change anymore."""
25
+ return self
26
+
27
+
28
+ def load_midas_transform(model_type):
29
+ # https://github.com/isl-org/MiDaS/blob/master/run.py
30
+ # load transform only
31
+ if model_type == "dpt_large": # DPT-Large
32
+ net_w, net_h = 384, 384
33
+ resize_mode = "minimal"
34
+ normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
35
+
36
+ elif model_type == "dpt_hybrid": # DPT-Hybrid
37
+ net_w, net_h = 384, 384
38
+ resize_mode = "minimal"
39
+ normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
40
+
41
+ elif model_type == "midas_v21":
42
+ net_w, net_h = 384, 384
43
+ resize_mode = "upper_bound"
44
+ normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
45
+
46
+ elif model_type == "midas_v21_small":
47
+ net_w, net_h = 256, 256
48
+ resize_mode = "upper_bound"
49
+ normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
50
+
51
+ else:
52
+ assert False, f"model_type '{model_type}' not implemented, use: --model_type large"
53
+
54
+ transform = Compose(
55
+ [
56
+ Resize(
57
+ net_w,
58
+ net_h,
59
+ resize_target=None,
60
+ keep_aspect_ratio=True,
61
+ ensure_multiple_of=32,
62
+ resize_method=resize_mode,
63
+ image_interpolation_method=cv2.INTER_CUBIC,
64
+ ),
65
+ normalization,
66
+ PrepareForNet(),
67
+ ]
68
+ )
69
+
70
+ return transform
71
+
72
+
73
+ def load_model(model_type):
74
+ # https://github.com/isl-org/MiDaS/blob/master/run.py
75
+ # load network
76
+ model_path = ISL_PATHS[model_type]
77
+ if model_type == "dpt_large": # DPT-Large
78
+ model = DPTDepthModel(
79
+ path=model_path,
80
+ backbone="vitl16_384",
81
+ non_negative=True,
82
+ )
83
+ net_w, net_h = 384, 384
84
+ resize_mode = "minimal"
85
+ normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
86
+
87
+ elif model_type == "dpt_hybrid": # DPT-Hybrid
88
+ model = DPTDepthModel(
89
+ path=model_path,
90
+ backbone="vitb_rn50_384",
91
+ non_negative=True,
92
+ )
93
+ net_w, net_h = 384, 384
94
+ resize_mode = "minimal"
95
+ normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
96
+
97
+ elif model_type == "midas_v21":
98
+ model = MidasNet(model_path, non_negative=True)
99
+ net_w, net_h = 384, 384
100
+ resize_mode = "upper_bound"
101
+ normalization = NormalizeImage(
102
+ mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
103
+ )
104
+
105
+ elif model_type == "midas_v21_small":
106
+ model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
107
+ non_negative=True, blocks={'expand': True})
108
+ net_w, net_h = 256, 256
109
+ resize_mode = "upper_bound"
110
+ normalization = NormalizeImage(
111
+ mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
112
+ )
113
+
114
+ else:
115
+ print(f"model_type '{model_type}' not implemented, use: --model_type large")
116
+ assert False
117
+
118
+ transform = Compose(
119
+ [
120
+ Resize(
121
+ net_w,
122
+ net_h,
123
+ resize_target=None,
124
+ keep_aspect_ratio=True,
125
+ ensure_multiple_of=32,
126
+ resize_method=resize_mode,
127
+ image_interpolation_method=cv2.INTER_CUBIC,
128
+ ),
129
+ normalization,
130
+ PrepareForNet(),
131
+ ]
132
+ )
133
+
134
+ return model.eval(), transform
135
+
136
+
137
+ class MiDaSInference(nn.Module):
138
+ MODEL_TYPES_TORCH_HUB = [
139
+ "DPT_Large",
140
+ "DPT_Hybrid",
141
+ "MiDaS_small"
142
+ ]
143
+ MODEL_TYPES_ISL = [
144
+ "dpt_large",
145
+ "dpt_hybrid",
146
+ "midas_v21",
147
+ "midas_v21_small",
148
+ ]
149
+
150
+ def __init__(self, model_type):
151
+ super().__init__()
152
+ assert (model_type in self.MODEL_TYPES_ISL)
153
+ model, _ = load_model(model_type)
154
+ self.model = model
155
+ self.model.train = disabled_train
156
+
157
+ def forward(self, x):
158
+ # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array
159
+ # NOTE: we expect that the correct transform has been called during dataloading.
160
+ with torch.no_grad():
161
+ prediction = self.model(x)
162
+ prediction = torch.nn.functional.interpolate(
163
+ prediction.unsqueeze(1),
164
+ size=x.shape[2:],
165
+ mode="bicubic",
166
+ align_corners=False,
167
+ )
168
+ assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3])
169
+ return prediction
170
+
ldm/modules/midas/midas/__init__.py ADDED
File without changes
ldm/modules/midas/midas/base_model.py ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+
4
+ class BaseModel(torch.nn.Module):
5
+ def load(self, path):
6
+ """Load model from file.
7
+
8
+ Args:
9
+ path (str): file path
10
+ """
11
+ parameters = torch.load(path, map_location=torch.device('cpu'))
12
+
13
+ if "optimizer" in parameters:
14
+ parameters = parameters["model"]
15
+
16
+ self.load_state_dict(parameters)
ldm/modules/midas/midas/blocks.py ADDED
@@ -0,0 +1,342 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+
4
+ from .vit import (
5
+ _make_pretrained_vitb_rn50_384,
6
+ _make_pretrained_vitl16_384,
7
+ _make_pretrained_vitb16_384,
8
+ forward_vit,
9
+ )
10
+
11
+ def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",):
12
+ if backbone == "vitl16_384":
13
+ pretrained = _make_pretrained_vitl16_384(
14
+ use_pretrained, hooks=hooks, use_readout=use_readout
15
+ )
16
+ scratch = _make_scratch(
17
+ [256, 512, 1024, 1024], features, groups=groups, expand=expand
18
+ ) # ViT-L/16 - 85.0% Top1 (backbone)
19
+ elif backbone == "vitb_rn50_384":
20
+ pretrained = _make_pretrained_vitb_rn50_384(
21
+ use_pretrained,
22
+ hooks=hooks,
23
+ use_vit_only=use_vit_only,
24
+ use_readout=use_readout,
25
+ )
26
+ scratch = _make_scratch(
27
+ [256, 512, 768, 768], features, groups=groups, expand=expand
28
+ ) # ViT-H/16 - 85.0% Top1 (backbone)
29
+ elif backbone == "vitb16_384":
30
+ pretrained = _make_pretrained_vitb16_384(
31
+ use_pretrained, hooks=hooks, use_readout=use_readout
32
+ )
33
+ scratch = _make_scratch(
34
+ [96, 192, 384, 768], features, groups=groups, expand=expand
35
+ ) # ViT-B/16 - 84.6% Top1 (backbone)
36
+ elif backbone == "resnext101_wsl":
37
+ pretrained = _make_pretrained_resnext101_wsl(use_pretrained)
38
+ scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand) # efficientnet_lite3
39
+ elif backbone == "efficientnet_lite3":
40
+ pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable)
41
+ scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand) # efficientnet_lite3
42
+ else:
43
+ print(f"Backbone '{backbone}' not implemented")
44
+ assert False
45
+
46
+ return pretrained, scratch
47
+
48
+
49
+ def _make_scratch(in_shape, out_shape, groups=1, expand=False):
50
+ scratch = nn.Module()
51
+
52
+ out_shape1 = out_shape
53
+ out_shape2 = out_shape
54
+ out_shape3 = out_shape
55
+ out_shape4 = out_shape
56
+ if expand==True:
57
+ out_shape1 = out_shape
58
+ out_shape2 = out_shape*2
59
+ out_shape3 = out_shape*4
60
+ out_shape4 = out_shape*8
61
+
62
+ scratch.layer1_rn = nn.Conv2d(
63
+ in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
64
+ )
65
+ scratch.layer2_rn = nn.Conv2d(
66
+ in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
67
+ )
68
+ scratch.layer3_rn = nn.Conv2d(
69
+ in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
70
+ )
71
+ scratch.layer4_rn = nn.Conv2d(
72
+ in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
73
+ )
74
+
75
+ return scratch
76
+
77
+
78
+ def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False):
79
+ efficientnet = torch.hub.load(
80
+ "rwightman/gen-efficientnet-pytorch",
81
+ "tf_efficientnet_lite3",
82
+ pretrained=use_pretrained,
83
+ exportable=exportable
84
+ )
85
+ return _make_efficientnet_backbone(efficientnet)
86
+
87
+
88
+ def _make_efficientnet_backbone(effnet):
89
+ pretrained = nn.Module()
90
+
91
+ pretrained.layer1 = nn.Sequential(
92
+ effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2]
93
+ )
94
+ pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3])
95
+ pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5])
96
+ pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9])
97
+
98
+ return pretrained
99
+
100
+
101
+ def _make_resnet_backbone(resnet):
102
+ pretrained = nn.Module()
103
+ pretrained.layer1 = nn.Sequential(
104
+ resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1
105
+ )
106
+
107
+ pretrained.layer2 = resnet.layer2
108
+ pretrained.layer3 = resnet.layer3
109
+ pretrained.layer4 = resnet.layer4
110
+
111
+ return pretrained
112
+
113
+
114
+ def _make_pretrained_resnext101_wsl(use_pretrained):
115
+ resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl")
116
+ return _make_resnet_backbone(resnet)
117
+
118
+
119
+
120
+ class Interpolate(nn.Module):
121
+ """Interpolation module.
122
+ """
123
+
124
+ def __init__(self, scale_factor, mode, align_corners=False):
125
+ """Init.
126
+
127
+ Args:
128
+ scale_factor (float): scaling
129
+ mode (str): interpolation mode
130
+ """
131
+ super(Interpolate, self).__init__()
132
+
133
+ self.interp = nn.functional.interpolate
134
+ self.scale_factor = scale_factor
135
+ self.mode = mode
136
+ self.align_corners = align_corners
137
+
138
+ def forward(self, x):
139
+ """Forward pass.
140
+
141
+ Args:
142
+ x (tensor): input
143
+
144
+ Returns:
145
+ tensor: interpolated data
146
+ """
147
+
148
+ x = self.interp(
149
+ x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners
150
+ )
151
+
152
+ return x
153
+
154
+
155
+ class ResidualConvUnit(nn.Module):
156
+ """Residual convolution module.
157
+ """
158
+
159
+ def __init__(self, features):
160
+ """Init.
161
+
162
+ Args:
163
+ features (int): number of features
164
+ """
165
+ super().__init__()
166
+
167
+ self.conv1 = nn.Conv2d(
168
+ features, features, kernel_size=3, stride=1, padding=1, bias=True
169
+ )
170
+
171
+ self.conv2 = nn.Conv2d(
172
+ features, features, kernel_size=3, stride=1, padding=1, bias=True
173
+ )
174
+
175
+ self.relu = nn.ReLU(inplace=True)
176
+
177
+ def forward(self, x):
178
+ """Forward pass.
179
+
180
+ Args:
181
+ x (tensor): input
182
+
183
+ Returns:
184
+ tensor: output
185
+ """
186
+ out = self.relu(x)
187
+ out = self.conv1(out)
188
+ out = self.relu(out)
189
+ out = self.conv2(out)
190
+
191
+ return out + x
192
+
193
+
194
+ class FeatureFusionBlock(nn.Module):
195
+ """Feature fusion block.
196
+ """
197
+
198
+ def __init__(self, features):
199
+ """Init.
200
+
201
+ Args:
202
+ features (int): number of features
203
+ """
204
+ super(FeatureFusionBlock, self).__init__()
205
+
206
+ self.resConfUnit1 = ResidualConvUnit(features)
207
+ self.resConfUnit2 = ResidualConvUnit(features)
208
+
209
+ def forward(self, *xs):
210
+ """Forward pass.
211
+
212
+ Returns:
213
+ tensor: output
214
+ """
215
+ output = xs[0]
216
+
217
+ if len(xs) == 2:
218
+ output += self.resConfUnit1(xs[1])
219
+
220
+ output = self.resConfUnit2(output)
221
+
222
+ output = nn.functional.interpolate(
223
+ output, scale_factor=2, mode="bilinear", align_corners=True
224
+ )
225
+
226
+ return output
227
+
228
+
229
+
230
+
231
+ class ResidualConvUnit_custom(nn.Module):
232
+ """Residual convolution module.
233
+ """
234
+
235
+ def __init__(self, features, activation, bn):
236
+ """Init.
237
+
238
+ Args:
239
+ features (int): number of features
240
+ """
241
+ super().__init__()
242
+
243
+ self.bn = bn
244
+
245
+ self.groups=1
246
+
247
+ self.conv1 = nn.Conv2d(
248
+ features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
249
+ )
250
+
251
+ self.conv2 = nn.Conv2d(
252
+ features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
253
+ )
254
+
255
+ if self.bn==True:
256
+ self.bn1 = nn.BatchNorm2d(features)
257
+ self.bn2 = nn.BatchNorm2d(features)
258
+
259
+ self.activation = activation
260
+
261
+ self.skip_add = nn.quantized.FloatFunctional()
262
+
263
+ def forward(self, x):
264
+ """Forward pass.
265
+
266
+ Args:
267
+ x (tensor): input
268
+
269
+ Returns:
270
+ tensor: output
271
+ """
272
+
273
+ out = self.activation(x)
274
+ out = self.conv1(out)
275
+ if self.bn==True:
276
+ out = self.bn1(out)
277
+
278
+ out = self.activation(out)
279
+ out = self.conv2(out)
280
+ if self.bn==True:
281
+ out = self.bn2(out)
282
+
283
+ if self.groups > 1:
284
+ out = self.conv_merge(out)
285
+
286
+ return self.skip_add.add(out, x)
287
+
288
+ # return out + x
289
+
290
+
291
+ class FeatureFusionBlock_custom(nn.Module):
292
+ """Feature fusion block.
293
+ """
294
+
295
+ def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True):
296
+ """Init.
297
+
298
+ Args:
299
+ features (int): number of features
300
+ """
301
+ super(FeatureFusionBlock_custom, self).__init__()
302
+
303
+ self.deconv = deconv
304
+ self.align_corners = align_corners
305
+
306
+ self.groups=1
307
+
308
+ self.expand = expand
309
+ out_features = features
310
+ if self.expand==True:
311
+ out_features = features//2
312
+
313
+ self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
314
+
315
+ self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)
316
+ self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)
317
+
318
+ self.skip_add = nn.quantized.FloatFunctional()
319
+
320
+ def forward(self, *xs):
321
+ """Forward pass.
322
+
323
+ Returns:
324
+ tensor: output
325
+ """
326
+ output = xs[0]
327
+
328
+ if len(xs) == 2:
329
+ res = self.resConfUnit1(xs[1])
330
+ output = self.skip_add.add(output, res)
331
+ # output += res
332
+
333
+ output = self.resConfUnit2(output)
334
+
335
+ output = nn.functional.interpolate(
336
+ output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
337
+ )
338
+
339
+ output = self.out_conv(output)
340
+
341
+ return output
342
+
ldm/modules/midas/midas/dpt_depth.py ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ import torch.nn.functional as F
4
+
5
+ from .base_model import BaseModel
6
+ from .blocks import (
7
+ FeatureFusionBlock,
8
+ FeatureFusionBlock_custom,
9
+ Interpolate,
10
+ _make_encoder,
11
+ forward_vit,
12
+ )
13
+
14
+
15
+ def _make_fusion_block(features, use_bn):
16
+ return FeatureFusionBlock_custom(
17
+ features,
18
+ nn.ReLU(False),
19
+ deconv=False,
20
+ bn=use_bn,
21
+ expand=False,
22
+ align_corners=True,
23
+ )
24
+
25
+
26
+ class DPT(BaseModel):
27
+ def __init__(
28
+ self,
29
+ head,
30
+ features=256,
31
+ backbone="vitb_rn50_384",
32
+ readout="project",
33
+ channels_last=False,
34
+ use_bn=False,
35
+ ):
36
+
37
+ super(DPT, self).__init__()
38
+
39
+ self.channels_last = channels_last
40
+
41
+ hooks = {
42
+ "vitb_rn50_384": [0, 1, 8, 11],
43
+ "vitb16_384": [2, 5, 8, 11],
44
+ "vitl16_384": [5, 11, 17, 23],
45
+ }
46
+
47
+ # Instantiate backbone and reassemble blocks
48
+ self.pretrained, self.scratch = _make_encoder(
49
+ backbone,
50
+ features,
51
+ False, # Set to true of you want to train from scratch, uses ImageNet weights
52
+ groups=1,
53
+ expand=False,
54
+ exportable=False,
55
+ hooks=hooks[backbone],
56
+ use_readout=readout,
57
+ )
58
+
59
+ self.scratch.refinenet1 = _make_fusion_block(features, use_bn)
60
+ self.scratch.refinenet2 = _make_fusion_block(features, use_bn)
61
+ self.scratch.refinenet3 = _make_fusion_block(features, use_bn)
62
+ self.scratch.refinenet4 = _make_fusion_block(features, use_bn)
63
+
64
+ self.scratch.output_conv = head
65
+
66
+
67
+ def forward(self, x):
68
+ if self.channels_last == True:
69
+ x.contiguous(memory_format=torch.channels_last)
70
+
71
+ layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x)
72
+
73
+ layer_1_rn = self.scratch.layer1_rn(layer_1)
74
+ layer_2_rn = self.scratch.layer2_rn(layer_2)
75
+ layer_3_rn = self.scratch.layer3_rn(layer_3)
76
+ layer_4_rn = self.scratch.layer4_rn(layer_4)
77
+
78
+ path_4 = self.scratch.refinenet4(layer_4_rn)
79
+ path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
80
+ path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
81
+ path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
82
+
83
+ out = self.scratch.output_conv(path_1)
84
+
85
+ return out
86
+
87
+
88
+ class DPTDepthModel(DPT):
89
+ def __init__(self, path=None, non_negative=True, **kwargs):
90
+ features = kwargs["features"] if "features" in kwargs else 256
91
+
92
+ head = nn.Sequential(
93
+ nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1),
94
+ Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
95
+ nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1),
96
+ nn.ReLU(True),
97
+ nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
98
+ nn.ReLU(True) if non_negative else nn.Identity(),
99
+ nn.Identity(),
100
+ )
101
+
102
+ super().__init__(head, **kwargs)
103
+
104
+ if path is not None:
105
+ self.load(path)
106
+
107
+ def forward(self, x):
108
+ return super().forward(x).squeeze(dim=1)
109
+
ldm/modules/midas/midas/midas_net.py ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """MidashNet: Network for monocular depth estimation trained by mixing several datasets.
2
+ This file contains code that is adapted from
3
+ https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
4
+ """
5
+ import torch
6
+ import torch.nn as nn
7
+
8
+ from .base_model import BaseModel
9
+ from .blocks import FeatureFusionBlock, Interpolate, _make_encoder
10
+
11
+
12
+ class MidasNet(BaseModel):
13
+ """Network for monocular depth estimation.
14
+ """
15
+
16
+ def __init__(self, path=None, features=256, non_negative=True):
17
+ """Init.
18
+
19
+ Args:
20
+ path (str, optional): Path to saved model. Defaults to None.
21
+ features (int, optional): Number of features. Defaults to 256.
22
+ backbone (str, optional): Backbone network for encoder. Defaults to resnet50
23
+ """
24
+ print("Loading weights: ", path)
25
+
26
+ super(MidasNet, self).__init__()
27
+
28
+ use_pretrained = False if path is None else True
29
+
30
+ self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained)
31
+
32
+ self.scratch.refinenet4 = FeatureFusionBlock(features)
33
+ self.scratch.refinenet3 = FeatureFusionBlock(features)
34
+ self.scratch.refinenet2 = FeatureFusionBlock(features)
35
+ self.scratch.refinenet1 = FeatureFusionBlock(features)
36
+
37
+ self.scratch.output_conv = nn.Sequential(
38
+ nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1),
39
+ Interpolate(scale_factor=2, mode="bilinear"),
40
+ nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1),
41
+ nn.ReLU(True),
42
+ nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
43
+ nn.ReLU(True) if non_negative else nn.Identity(),
44
+ )
45
+
46
+ if path:
47
+ self.load(path)
48
+
49
+ def forward(self, x):
50
+ """Forward pass.
51
+
52
+ Args:
53
+ x (tensor): input data (image)
54
+
55
+ Returns:
56
+ tensor: depth
57
+ """
58
+
59
+ layer_1 = self.pretrained.layer1(x)
60
+ layer_2 = self.pretrained.layer2(layer_1)
61
+ layer_3 = self.pretrained.layer3(layer_2)
62
+ layer_4 = self.pretrained.layer4(layer_3)
63
+
64
+ layer_1_rn = self.scratch.layer1_rn(layer_1)
65
+ layer_2_rn = self.scratch.layer2_rn(layer_2)
66
+ layer_3_rn = self.scratch.layer3_rn(layer_3)
67
+ layer_4_rn = self.scratch.layer4_rn(layer_4)
68
+
69
+ path_4 = self.scratch.refinenet4(layer_4_rn)
70
+ path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
71
+ path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
72
+ path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
73
+
74
+ out = self.scratch.output_conv(path_1)
75
+
76
+ return torch.squeeze(out, dim=1)
ldm/modules/midas/midas/midas_net_custom.py ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """MidashNet: Network for monocular depth estimation trained by mixing several datasets.
2
+ This file contains code that is adapted from
3
+ https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
4
+ """
5
+ import torch
6
+ import torch.nn as nn
7
+
8
+ from .base_model import BaseModel
9
+ from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder
10
+
11
+
12
+ class MidasNet_small(BaseModel):
13
+ """Network for monocular depth estimation.
14
+ """
15
+
16
+ def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True,
17
+ blocks={'expand': True}):
18
+ """Init.
19
+
20
+ Args:
21
+ path (str, optional): Path to saved model. Defaults to None.
22
+ features (int, optional): Number of features. Defaults to 256.
23
+ backbone (str, optional): Backbone network for encoder. Defaults to resnet50
24
+ """
25
+ print("Loading weights: ", path)
26
+
27
+ super(MidasNet_small, self).__init__()
28
+
29
+ use_pretrained = False if path else True
30
+
31
+ self.channels_last = channels_last
32
+ self.blocks = blocks
33
+ self.backbone = backbone
34
+
35
+ self.groups = 1
36
+
37
+ features1=features
38
+ features2=features
39
+ features3=features
40
+ features4=features
41
+ self.expand = False
42
+ if "expand" in self.blocks and self.blocks['expand'] == True:
43
+ self.expand = True
44
+ features1=features
45
+ features2=features*2
46
+ features3=features*4
47
+ features4=features*8
48
+
49
+ self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable)
50
+
51
+ self.scratch.activation = nn.ReLU(False)
52
+
53
+ self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
54
+ self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
55
+ self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
56
+ self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners)
57
+
58
+
59
+ self.scratch.output_conv = nn.Sequential(
60
+ nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups),
61
+ Interpolate(scale_factor=2, mode="bilinear"),
62
+ nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1),
63
+ self.scratch.activation,
64
+ nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
65
+ nn.ReLU(True) if non_negative else nn.Identity(),
66
+ nn.Identity(),
67
+ )
68
+
69
+ if path:
70
+ self.load(path)
71
+
72
+
73
+ def forward(self, x):
74
+ """Forward pass.
75
+
76
+ Args:
77
+ x (tensor): input data (image)
78
+
79
+ Returns:
80
+ tensor: depth
81
+ """
82
+ if self.channels_last==True:
83
+ print("self.channels_last = ", self.channels_last)
84
+ x.contiguous(memory_format=torch.channels_last)
85
+
86
+
87
+ layer_1 = self.pretrained.layer1(x)
88
+ layer_2 = self.pretrained.layer2(layer_1)
89
+ layer_3 = self.pretrained.layer3(layer_2)
90
+ layer_4 = self.pretrained.layer4(layer_3)
91
+
92
+ layer_1_rn = self.scratch.layer1_rn(layer_1)
93
+ layer_2_rn = self.scratch.layer2_rn(layer_2)
94
+ layer_3_rn = self.scratch.layer3_rn(layer_3)
95
+ layer_4_rn = self.scratch.layer4_rn(layer_4)
96
+
97
+
98
+ path_4 = self.scratch.refinenet4(layer_4_rn)
99
+ path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
100
+ path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
101
+ path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
102
+
103
+ out = self.scratch.output_conv(path_1)
104
+
105
+ return torch.squeeze(out, dim=1)
106
+
107
+
108
+
109
+ def fuse_model(m):
110
+ prev_previous_type = nn.Identity()
111
+ prev_previous_name = ''
112
+ previous_type = nn.Identity()
113
+ previous_name = ''
114
+ for name, module in m.named_modules():
115
+ if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU:
116
+ # print("FUSED ", prev_previous_name, previous_name, name)
117
+ torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True)
118
+ elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d:
119
+ # print("FUSED ", prev_previous_name, previous_name)
120
+ torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True)
121
+ # elif previous_type == nn.Conv2d and type(module) == nn.ReLU:
122
+ # print("FUSED ", previous_name, name)
123
+ # torch.quantization.fuse_modules(m, [previous_name, name], inplace=True)
124
+
125
+ prev_previous_type = previous_type
126
+ prev_previous_name = previous_name
127
+ previous_type = type(module)
128
+ previous_name = name
ldm/modules/midas/midas/transforms.py ADDED
@@ -0,0 +1,234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import cv2
3
+ import math
4
+
5
+
6
+ def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA):
7
+ """Rezise the sample to ensure the given size. Keeps aspect ratio.
8
+
9
+ Args:
10
+ sample (dict): sample
11
+ size (tuple): image size
12
+
13
+ Returns:
14
+ tuple: new size
15
+ """
16
+ shape = list(sample["disparity"].shape)
17
+
18
+ if shape[0] >= size[0] and shape[1] >= size[1]:
19
+ return sample
20
+
21
+ scale = [0, 0]
22
+ scale[0] = size[0] / shape[0]
23
+ scale[1] = size[1] / shape[1]
24
+
25
+ scale = max(scale)
26
+
27
+ shape[0] = math.ceil(scale * shape[0])
28
+ shape[1] = math.ceil(scale * shape[1])
29
+
30
+ # resize
31
+ sample["image"] = cv2.resize(
32
+ sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method
33
+ )
34
+
35
+ sample["disparity"] = cv2.resize(
36
+ sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST
37
+ )
38
+ sample["mask"] = cv2.resize(
39
+ sample["mask"].astype(np.float32),
40
+ tuple(shape[::-1]),
41
+ interpolation=cv2.INTER_NEAREST,
42
+ )
43
+ sample["mask"] = sample["mask"].astype(bool)
44
+
45
+ return tuple(shape)
46
+
47
+
48
+ class Resize(object):
49
+ """Resize sample to given size (width, height).
50
+ """
51
+
52
+ def __init__(
53
+ self,
54
+ width,
55
+ height,
56
+ resize_target=True,
57
+ keep_aspect_ratio=False,
58
+ ensure_multiple_of=1,
59
+ resize_method="lower_bound",
60
+ image_interpolation_method=cv2.INTER_AREA,
61
+ ):
62
+ """Init.
63
+
64
+ Args:
65
+ width (int): desired output width
66
+ height (int): desired output height
67
+ resize_target (bool, optional):
68
+ True: Resize the full sample (image, mask, target).
69
+ False: Resize image only.
70
+ Defaults to True.
71
+ keep_aspect_ratio (bool, optional):
72
+ True: Keep the aspect ratio of the input sample.
73
+ Output sample might not have the given width and height, and
74
+ resize behaviour depends on the parameter 'resize_method'.
75
+ Defaults to False.
76
+ ensure_multiple_of (int, optional):
77
+ Output width and height is constrained to be multiple of this parameter.
78
+ Defaults to 1.
79
+ resize_method (str, optional):
80
+ "lower_bound": Output will be at least as large as the given size.
81
+ "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.)
82
+ "minimal": Scale as least as possible. (Output size might be smaller than given size.)
83
+ Defaults to "lower_bound".
84
+ """
85
+ self.__width = width
86
+ self.__height = height
87
+
88
+ self.__resize_target = resize_target
89
+ self.__keep_aspect_ratio = keep_aspect_ratio
90
+ self.__multiple_of = ensure_multiple_of
91
+ self.__resize_method = resize_method
92
+ self.__image_interpolation_method = image_interpolation_method
93
+
94
+ def constrain_to_multiple_of(self, x, min_val=0, max_val=None):
95
+ y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int)
96
+
97
+ if max_val is not None and y > max_val:
98
+ y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int)
99
+
100
+ if y < min_val:
101
+ y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int)
102
+
103
+ return y
104
+
105
+ def get_size(self, width, height):
106
+ # determine new height and width
107
+ scale_height = self.__height / height
108
+ scale_width = self.__width / width
109
+
110
+ if self.__keep_aspect_ratio:
111
+ if self.__resize_method == "lower_bound":
112
+ # scale such that output size is lower bound
113
+ if scale_width > scale_height:
114
+ # fit width
115
+ scale_height = scale_width
116
+ else:
117
+ # fit height
118
+ scale_width = scale_height
119
+ elif self.__resize_method == "upper_bound":
120
+ # scale such that output size is upper bound
121
+ if scale_width < scale_height:
122
+ # fit width
123
+ scale_height = scale_width
124
+ else:
125
+ # fit height
126
+ scale_width = scale_height
127
+ elif self.__resize_method == "minimal":
128
+ # scale as least as possbile
129
+ if abs(1 - scale_width) < abs(1 - scale_height):
130
+ # fit width
131
+ scale_height = scale_width
132
+ else:
133
+ # fit height
134
+ scale_width = scale_height
135
+ else:
136
+ raise ValueError(
137
+ f"resize_method {self.__resize_method} not implemented"
138
+ )
139
+
140
+ if self.__resize_method == "lower_bound":
141
+ new_height = self.constrain_to_multiple_of(
142
+ scale_height * height, min_val=self.__height
143
+ )
144
+ new_width = self.constrain_to_multiple_of(
145
+ scale_width * width, min_val=self.__width
146
+ )
147
+ elif self.__resize_method == "upper_bound":
148
+ new_height = self.constrain_to_multiple_of(
149
+ scale_height * height, max_val=self.__height
150
+ )
151
+ new_width = self.constrain_to_multiple_of(
152
+ scale_width * width, max_val=self.__width
153
+ )
154
+ elif self.__resize_method == "minimal":
155
+ new_height = self.constrain_to_multiple_of(scale_height * height)
156
+ new_width = self.constrain_to_multiple_of(scale_width * width)
157
+ else:
158
+ raise ValueError(f"resize_method {self.__resize_method} not implemented")
159
+
160
+ return (new_width, new_height)
161
+
162
+ def __call__(self, sample):
163
+ width, height = self.get_size(
164
+ sample["image"].shape[1], sample["image"].shape[0]
165
+ )
166
+
167
+ # resize sample
168
+ sample["image"] = cv2.resize(
169
+ sample["image"],
170
+ (width, height),
171
+ interpolation=self.__image_interpolation_method,
172
+ )
173
+
174
+ if self.__resize_target:
175
+ if "disparity" in sample:
176
+ sample["disparity"] = cv2.resize(
177
+ sample["disparity"],
178
+ (width, height),
179
+ interpolation=cv2.INTER_NEAREST,
180
+ )
181
+
182
+ if "depth" in sample:
183
+ sample["depth"] = cv2.resize(
184
+ sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST
185
+ )
186
+
187
+ sample["mask"] = cv2.resize(
188
+ sample["mask"].astype(np.float32),
189
+ (width, height),
190
+ interpolation=cv2.INTER_NEAREST,
191
+ )
192
+ sample["mask"] = sample["mask"].astype(bool)
193
+
194
+ return sample
195
+
196
+
197
+ class NormalizeImage(object):
198
+ """Normlize image by given mean and std.
199
+ """
200
+
201
+ def __init__(self, mean, std):
202
+ self.__mean = mean
203
+ self.__std = std
204
+
205
+ def __call__(self, sample):
206
+ sample["image"] = (sample["image"] - self.__mean) / self.__std
207
+
208
+ return sample
209
+
210
+
211
+ class PrepareForNet(object):
212
+ """Prepare sample for usage as network input.
213
+ """
214
+
215
+ def __init__(self):
216
+ pass
217
+
218
+ def __call__(self, sample):
219
+ image = np.transpose(sample["image"], (2, 0, 1))
220
+ sample["image"] = np.ascontiguousarray(image).astype(np.float32)
221
+
222
+ if "mask" in sample:
223
+ sample["mask"] = sample["mask"].astype(np.float32)
224
+ sample["mask"] = np.ascontiguousarray(sample["mask"])
225
+
226
+ if "disparity" in sample:
227
+ disparity = sample["disparity"].astype(np.float32)
228
+ sample["disparity"] = np.ascontiguousarray(disparity)
229
+
230
+ if "depth" in sample:
231
+ depth = sample["depth"].astype(np.float32)
232
+ sample["depth"] = np.ascontiguousarray(depth)
233
+
234
+ return sample
ldm/modules/midas/midas/vit.py ADDED
@@ -0,0 +1,491 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ import timm
4
+ import types
5
+ import math
6
+ import torch.nn.functional as F
7
+
8
+
9
+ class Slice(nn.Module):
10
+ def __init__(self, start_index=1):
11
+ super(Slice, self).__init__()
12
+ self.start_index = start_index
13
+
14
+ def forward(self, x):
15
+ return x[:, self.start_index :]
16
+
17
+
18
+ class AddReadout(nn.Module):
19
+ def __init__(self, start_index=1):
20
+ super(AddReadout, self).__init__()
21
+ self.start_index = start_index
22
+
23
+ def forward(self, x):
24
+ if self.start_index == 2:
25
+ readout = (x[:, 0] + x[:, 1]) / 2
26
+ else:
27
+ readout = x[:, 0]
28
+ return x[:, self.start_index :] + readout.unsqueeze(1)
29
+
30
+
31
+ class ProjectReadout(nn.Module):
32
+ def __init__(self, in_features, start_index=1):
33
+ super(ProjectReadout, self).__init__()
34
+ self.start_index = start_index
35
+
36
+ self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU())
37
+
38
+ def forward(self, x):
39
+ readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :])
40
+ features = torch.cat((x[:, self.start_index :], readout), -1)
41
+
42
+ return self.project(features)
43
+
44
+
45
+ class Transpose(nn.Module):
46
+ def __init__(self, dim0, dim1):
47
+ super(Transpose, self).__init__()
48
+ self.dim0 = dim0
49
+ self.dim1 = dim1
50
+
51
+ def forward(self, x):
52
+ x = x.transpose(self.dim0, self.dim1)
53
+ return x
54
+
55
+
56
+ def forward_vit(pretrained, x):
57
+ b, c, h, w = x.shape
58
+
59
+ glob = pretrained.model.forward_flex(x)
60
+
61
+ layer_1 = pretrained.activations["1"]
62
+ layer_2 = pretrained.activations["2"]
63
+ layer_3 = pretrained.activations["3"]
64
+ layer_4 = pretrained.activations["4"]
65
+
66
+ layer_1 = pretrained.act_postprocess1[0:2](layer_1)
67
+ layer_2 = pretrained.act_postprocess2[0:2](layer_2)
68
+ layer_3 = pretrained.act_postprocess3[0:2](layer_3)
69
+ layer_4 = pretrained.act_postprocess4[0:2](layer_4)
70
+
71
+ unflatten = nn.Sequential(
72
+ nn.Unflatten(
73
+ 2,
74
+ torch.Size(
75
+ [
76
+ h // pretrained.model.patch_size[1],
77
+ w // pretrained.model.patch_size[0],
78
+ ]
79
+ ),
80
+ )
81
+ )
82
+
83
+ if layer_1.ndim == 3:
84
+ layer_1 = unflatten(layer_1)
85
+ if layer_2.ndim == 3:
86
+ layer_2 = unflatten(layer_2)
87
+ if layer_3.ndim == 3:
88
+ layer_3 = unflatten(layer_3)
89
+ if layer_4.ndim == 3:
90
+ layer_4 = unflatten(layer_4)
91
+
92
+ layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1)
93
+ layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2)
94
+ layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3)
95
+ layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4)
96
+
97
+ return layer_1, layer_2, layer_3, layer_4
98
+
99
+
100
+ def _resize_pos_embed(self, posemb, gs_h, gs_w):
101
+ posemb_tok, posemb_grid = (
102
+ posemb[:, : self.start_index],
103
+ posemb[0, self.start_index :],
104
+ )
105
+
106
+ gs_old = int(math.sqrt(len(posemb_grid)))
107
+
108
+ posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
109
+ posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear")
110
+ posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1)
111
+
112
+ posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
113
+
114
+ return posemb
115
+
116
+
117
+ def forward_flex(self, x):
118
+ b, c, h, w = x.shape
119
+
120
+ pos_embed = self._resize_pos_embed(
121
+ self.pos_embed, h // self.patch_size[1], w // self.patch_size[0]
122
+ )
123
+
124
+ B = x.shape[0]
125
+
126
+ if hasattr(self.patch_embed, "backbone"):
127
+ x = self.patch_embed.backbone(x)
128
+ if isinstance(x, (list, tuple)):
129
+ x = x[-1] # last feature if backbone outputs list/tuple of features
130
+
131
+ x = self.patch_embed.proj(x).flatten(2).transpose(1, 2)
132
+
133
+ if getattr(self, "dist_token", None) is not None:
134
+ cls_tokens = self.cls_token.expand(
135
+ B, -1, -1
136
+ ) # stole cls_tokens impl from Phil Wang, thanks
137
+ dist_token = self.dist_token.expand(B, -1, -1)
138
+ x = torch.cat((cls_tokens, dist_token, x), dim=1)
139
+ else:
140
+ cls_tokens = self.cls_token.expand(
141
+ B, -1, -1
142
+ ) # stole cls_tokens impl from Phil Wang, thanks
143
+ x = torch.cat((cls_tokens, x), dim=1)
144
+
145
+ x = x + pos_embed
146
+ x = self.pos_drop(x)
147
+
148
+ for blk in self.blocks:
149
+ x = blk(x)
150
+
151
+ x = self.norm(x)
152
+
153
+ return x
154
+
155
+
156
+ activations = {}
157
+
158
+
159
+ def get_activation(name):
160
+ def hook(model, input, output):
161
+ activations[name] = output
162
+
163
+ return hook
164
+
165
+
166
+ def get_readout_oper(vit_features, features, use_readout, start_index=1):
167
+ if use_readout == "ignore":
168
+ readout_oper = [Slice(start_index)] * len(features)
169
+ elif use_readout == "add":
170
+ readout_oper = [AddReadout(start_index)] * len(features)
171
+ elif use_readout == "project":
172
+ readout_oper = [
173
+ ProjectReadout(vit_features, start_index) for out_feat in features
174
+ ]
175
+ else:
176
+ assert (
177
+ False
178
+ ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'"
179
+
180
+ return readout_oper
181
+
182
+
183
+ def _make_vit_b16_backbone(
184
+ model,
185
+ features=[96, 192, 384, 768],
186
+ size=[384, 384],
187
+ hooks=[2, 5, 8, 11],
188
+ vit_features=768,
189
+ use_readout="ignore",
190
+ start_index=1,
191
+ ):
192
+ pretrained = nn.Module()
193
+
194
+ pretrained.model = model
195
+ pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
196
+ pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
197
+ pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
198
+ pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
199
+
200
+ pretrained.activations = activations
201
+
202
+ readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
203
+
204
+ # 32, 48, 136, 384
205
+ pretrained.act_postprocess1 = nn.Sequential(
206
+ readout_oper[0],
207
+ Transpose(1, 2),
208
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
209
+ nn.Conv2d(
210
+ in_channels=vit_features,
211
+ out_channels=features[0],
212
+ kernel_size=1,
213
+ stride=1,
214
+ padding=0,
215
+ ),
216
+ nn.ConvTranspose2d(
217
+ in_channels=features[0],
218
+ out_channels=features[0],
219
+ kernel_size=4,
220
+ stride=4,
221
+ padding=0,
222
+ bias=True,
223
+ dilation=1,
224
+ groups=1,
225
+ ),
226
+ )
227
+
228
+ pretrained.act_postprocess2 = nn.Sequential(
229
+ readout_oper[1],
230
+ Transpose(1, 2),
231
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
232
+ nn.Conv2d(
233
+ in_channels=vit_features,
234
+ out_channels=features[1],
235
+ kernel_size=1,
236
+ stride=1,
237
+ padding=0,
238
+ ),
239
+ nn.ConvTranspose2d(
240
+ in_channels=features[1],
241
+ out_channels=features[1],
242
+ kernel_size=2,
243
+ stride=2,
244
+ padding=0,
245
+ bias=True,
246
+ dilation=1,
247
+ groups=1,
248
+ ),
249
+ )
250
+
251
+ pretrained.act_postprocess3 = nn.Sequential(
252
+ readout_oper[2],
253
+ Transpose(1, 2),
254
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
255
+ nn.Conv2d(
256
+ in_channels=vit_features,
257
+ out_channels=features[2],
258
+ kernel_size=1,
259
+ stride=1,
260
+ padding=0,
261
+ ),
262
+ )
263
+
264
+ pretrained.act_postprocess4 = nn.Sequential(
265
+ readout_oper[3],
266
+ Transpose(1, 2),
267
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
268
+ nn.Conv2d(
269
+ in_channels=vit_features,
270
+ out_channels=features[3],
271
+ kernel_size=1,
272
+ stride=1,
273
+ padding=0,
274
+ ),
275
+ nn.Conv2d(
276
+ in_channels=features[3],
277
+ out_channels=features[3],
278
+ kernel_size=3,
279
+ stride=2,
280
+ padding=1,
281
+ ),
282
+ )
283
+
284
+ pretrained.model.start_index = start_index
285
+ pretrained.model.patch_size = [16, 16]
286
+
287
+ # We inject this function into the VisionTransformer instances so that
288
+ # we can use it with interpolated position embeddings without modifying the library source.
289
+ pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
290
+ pretrained.model._resize_pos_embed = types.MethodType(
291
+ _resize_pos_embed, pretrained.model
292
+ )
293
+
294
+ return pretrained
295
+
296
+
297
+ def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None):
298
+ model = timm.create_model("vit_large_patch16_384", pretrained=pretrained)
299
+
300
+ hooks = [5, 11, 17, 23] if hooks == None else hooks
301
+ return _make_vit_b16_backbone(
302
+ model,
303
+ features=[256, 512, 1024, 1024],
304
+ hooks=hooks,
305
+ vit_features=1024,
306
+ use_readout=use_readout,
307
+ )
308
+
309
+
310
+ def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None):
311
+ model = timm.create_model("vit_base_patch16_384", pretrained=pretrained)
312
+
313
+ hooks = [2, 5, 8, 11] if hooks == None else hooks
314
+ return _make_vit_b16_backbone(
315
+ model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
316
+ )
317
+
318
+
319
+ def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None):
320
+ model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained)
321
+
322
+ hooks = [2, 5, 8, 11] if hooks == None else hooks
323
+ return _make_vit_b16_backbone(
324
+ model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
325
+ )
326
+
327
+
328
+ def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None):
329
+ model = timm.create_model(
330
+ "vit_deit_base_distilled_patch16_384", pretrained=pretrained
331
+ )
332
+
333
+ hooks = [2, 5, 8, 11] if hooks == None else hooks
334
+ return _make_vit_b16_backbone(
335
+ model,
336
+ features=[96, 192, 384, 768],
337
+ hooks=hooks,
338
+ use_readout=use_readout,
339
+ start_index=2,
340
+ )
341
+
342
+
343
+ def _make_vit_b_rn50_backbone(
344
+ model,
345
+ features=[256, 512, 768, 768],
346
+ size=[384, 384],
347
+ hooks=[0, 1, 8, 11],
348
+ vit_features=768,
349
+ use_vit_only=False,
350
+ use_readout="ignore",
351
+ start_index=1,
352
+ ):
353
+ pretrained = nn.Module()
354
+
355
+ pretrained.model = model
356
+
357
+ if use_vit_only == True:
358
+ pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
359
+ pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
360
+ else:
361
+ pretrained.model.patch_embed.backbone.stages[0].register_forward_hook(
362
+ get_activation("1")
363
+ )
364
+ pretrained.model.patch_embed.backbone.stages[1].register_forward_hook(
365
+ get_activation("2")
366
+ )
367
+
368
+ pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
369
+ pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
370
+
371
+ pretrained.activations = activations
372
+
373
+ readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
374
+
375
+ if use_vit_only == True:
376
+ pretrained.act_postprocess1 = nn.Sequential(
377
+ readout_oper[0],
378
+ Transpose(1, 2),
379
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
380
+ nn.Conv2d(
381
+ in_channels=vit_features,
382
+ out_channels=features[0],
383
+ kernel_size=1,
384
+ stride=1,
385
+ padding=0,
386
+ ),
387
+ nn.ConvTranspose2d(
388
+ in_channels=features[0],
389
+ out_channels=features[0],
390
+ kernel_size=4,
391
+ stride=4,
392
+ padding=0,
393
+ bias=True,
394
+ dilation=1,
395
+ groups=1,
396
+ ),
397
+ )
398
+
399
+ pretrained.act_postprocess2 = nn.Sequential(
400
+ readout_oper[1],
401
+ Transpose(1, 2),
402
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
403
+ nn.Conv2d(
404
+ in_channels=vit_features,
405
+ out_channels=features[1],
406
+ kernel_size=1,
407
+ stride=1,
408
+ padding=0,
409
+ ),
410
+ nn.ConvTranspose2d(
411
+ in_channels=features[1],
412
+ out_channels=features[1],
413
+ kernel_size=2,
414
+ stride=2,
415
+ padding=0,
416
+ bias=True,
417
+ dilation=1,
418
+ groups=1,
419
+ ),
420
+ )
421
+ else:
422
+ pretrained.act_postprocess1 = nn.Sequential(
423
+ nn.Identity(), nn.Identity(), nn.Identity()
424
+ )
425
+ pretrained.act_postprocess2 = nn.Sequential(
426
+ nn.Identity(), nn.Identity(), nn.Identity()
427
+ )
428
+
429
+ pretrained.act_postprocess3 = nn.Sequential(
430
+ readout_oper[2],
431
+ Transpose(1, 2),
432
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
433
+ nn.Conv2d(
434
+ in_channels=vit_features,
435
+ out_channels=features[2],
436
+ kernel_size=1,
437
+ stride=1,
438
+ padding=0,
439
+ ),
440
+ )
441
+
442
+ pretrained.act_postprocess4 = nn.Sequential(
443
+ readout_oper[3],
444
+ Transpose(1, 2),
445
+ nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
446
+ nn.Conv2d(
447
+ in_channels=vit_features,
448
+ out_channels=features[3],
449
+ kernel_size=1,
450
+ stride=1,
451
+ padding=0,
452
+ ),
453
+ nn.Conv2d(
454
+ in_channels=features[3],
455
+ out_channels=features[3],
456
+ kernel_size=3,
457
+ stride=2,
458
+ padding=1,
459
+ ),
460
+ )
461
+
462
+ pretrained.model.start_index = start_index
463
+ pretrained.model.patch_size = [16, 16]
464
+
465
+ # We inject this function into the VisionTransformer instances so that
466
+ # we can use it with interpolated position embeddings without modifying the library source.
467
+ pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
468
+
469
+ # We inject this function into the VisionTransformer instances so that
470
+ # we can use it with interpolated position embeddings without modifying the library source.
471
+ pretrained.model._resize_pos_embed = types.MethodType(
472
+ _resize_pos_embed, pretrained.model
473
+ )
474
+
475
+ return pretrained
476
+
477
+
478
+ def _make_pretrained_vitb_rn50_384(
479
+ pretrained, use_readout="ignore", hooks=None, use_vit_only=False
480
+ ):
481
+ model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained)
482
+
483
+ hooks = [0, 1, 8, 11] if hooks == None else hooks
484
+ return _make_vit_b_rn50_backbone(
485
+ model,
486
+ features=[256, 512, 768, 768],
487
+ size=[384, 384],
488
+ hooks=hooks,
489
+ use_vit_only=use_vit_only,
490
+ use_readout=use_readout,
491
+ )
ldm/modules/midas/utils.py ADDED
@@ -0,0 +1,189 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Utils for monoDepth."""
2
+ import sys
3
+ import re
4
+ import numpy as np
5
+ import cv2
6
+ import torch
7
+
8
+
9
+ def read_pfm(path):
10
+ """Read pfm file.
11
+
12
+ Args:
13
+ path (str): path to file
14
+
15
+ Returns:
16
+ tuple: (data, scale)
17
+ """
18
+ with open(path, "rb") as file:
19
+
20
+ color = None
21
+ width = None
22
+ height = None
23
+ scale = None
24
+ endian = None
25
+
26
+ header = file.readline().rstrip()
27
+ if header.decode("ascii") == "PF":
28
+ color = True
29
+ elif header.decode("ascii") == "Pf":
30
+ color = False
31
+ else:
32
+ raise Exception("Not a PFM file: " + path)
33
+
34
+ dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii"))
35
+ if dim_match:
36
+ width, height = list(map(int, dim_match.groups()))
37
+ else:
38
+ raise Exception("Malformed PFM header.")
39
+
40
+ scale = float(file.readline().decode("ascii").rstrip())
41
+ if scale < 0:
42
+ # little-endian
43
+ endian = "<"
44
+ scale = -scale
45
+ else:
46
+ # big-endian
47
+ endian = ">"
48
+
49
+ data = np.fromfile(file, endian + "f")
50
+ shape = (height, width, 3) if color else (height, width)
51
+
52
+ data = np.reshape(data, shape)
53
+ data = np.flipud(data)
54
+
55
+ return data, scale
56
+
57
+
58
+ def write_pfm(path, image, scale=1):
59
+ """Write pfm file.
60
+
61
+ Args:
62
+ path (str): pathto file
63
+ image (array): data
64
+ scale (int, optional): Scale. Defaults to 1.
65
+ """
66
+
67
+ with open(path, "wb") as file:
68
+ color = None
69
+
70
+ if image.dtype.name != "float32":
71
+ raise Exception("Image dtype must be float32.")
72
+
73
+ image = np.flipud(image)
74
+
75
+ if len(image.shape) == 3 and image.shape[2] == 3: # color image
76
+ color = True
77
+ elif (
78
+ len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1
79
+ ): # greyscale
80
+ color = False
81
+ else:
82
+ raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.")
83
+
84
+ file.write("PF\n" if color else "Pf\n".encode())
85
+ file.write("%d %d\n".encode() % (image.shape[1], image.shape[0]))
86
+
87
+ endian = image.dtype.byteorder
88
+
89
+ if endian == "<" or endian == "=" and sys.byteorder == "little":
90
+ scale = -scale
91
+
92
+ file.write("%f\n".encode() % scale)
93
+
94
+ image.tofile(file)
95
+
96
+
97
+ def read_image(path):
98
+ """Read image and output RGB image (0-1).
99
+
100
+ Args:
101
+ path (str): path to file
102
+
103
+ Returns:
104
+ array: RGB image (0-1)
105
+ """
106
+ img = cv2.imread(path)
107
+
108
+ if img.ndim == 2:
109
+ img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
110
+
111
+ img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0
112
+
113
+ return img
114
+
115
+
116
+ def resize_image(img):
117
+ """Resize image and make it fit for network.
118
+
119
+ Args:
120
+ img (array): image
121
+
122
+ Returns:
123
+ tensor: data ready for network
124
+ """
125
+ height_orig = img.shape[0]
126
+ width_orig = img.shape[1]
127
+
128
+ if width_orig > height_orig:
129
+ scale = width_orig / 384
130
+ else:
131
+ scale = height_orig / 384
132
+
133
+ height = (np.ceil(height_orig / scale / 32) * 32).astype(int)
134
+ width = (np.ceil(width_orig / scale / 32) * 32).astype(int)
135
+
136
+ img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA)
137
+
138
+ img_resized = (
139
+ torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float()
140
+ )
141
+ img_resized = img_resized.unsqueeze(0)
142
+
143
+ return img_resized
144
+
145
+
146
+ def resize_depth(depth, width, height):
147
+ """Resize depth map and bring to CPU (numpy).
148
+
149
+ Args:
150
+ depth (tensor): depth
151
+ width (int): image width
152
+ height (int): image height
153
+
154
+ Returns:
155
+ array: processed depth
156
+ """
157
+ depth = torch.squeeze(depth[0, :, :, :]).to("cpu")
158
+
159
+ depth_resized = cv2.resize(
160
+ depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC
161
+ )
162
+
163
+ return depth_resized
164
+
165
+ def write_depth(path, depth, bits=1):
166
+ """Write depth map to pfm and png file.
167
+
168
+ Args:
169
+ path (str): filepath without extension
170
+ depth (array): depth
171
+ """
172
+ write_pfm(path + ".pfm", depth.astype(np.float32))
173
+
174
+ depth_min = depth.min()
175
+ depth_max = depth.max()
176
+
177
+ max_val = (2**(8*bits))-1
178
+
179
+ if depth_max - depth_min > np.finfo("float").eps:
180
+ out = max_val * (depth - depth_min) / (depth_max - depth_min)
181
+ else:
182
+ out = np.zeros(depth.shape, dtype=depth.type)
183
+
184
+ if bits == 1:
185
+ cv2.imwrite(path + ".png", out.astype("uint8"))
186
+ elif bits == 2:
187
+ cv2.imwrite(path + ".png", out.astype("uint16"))
188
+
189
+ return
ldm/modules/util.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+
4
+
5
+ def count_params(model):
6
+ total_params = sum(p.numel() for p in model.parameters())
7
+ return total_params
8
+
9
+
10
+ class ActNorm(nn.Module):
11
+ def __init__(self, num_features, logdet=False, affine=True,
12
+ allow_reverse_init=False):
13
+ assert affine
14
+ super().__init__()
15
+ self.logdet = logdet
16
+ self.loc = nn.Parameter(torch.zeros(1, num_features, 1, 1))
17
+ self.scale = nn.Parameter(torch.ones(1, num_features, 1, 1))
18
+ self.allow_reverse_init = allow_reverse_init
19
+
20
+ self.register_buffer('initialized', torch.tensor(0, dtype=torch.uint8))
21
+
22
+ def initialize(self, input):
23
+ with torch.no_grad():
24
+ flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1)
25
+ mean = (
26
+ flatten.mean(1)
27
+ .unsqueeze(1)
28
+ .unsqueeze(2)
29
+ .unsqueeze(3)
30
+ .permute(1, 0, 2, 3)
31
+ )
32
+ std = (
33
+ flatten.std(1)
34
+ .unsqueeze(1)
35
+ .unsqueeze(2)
36
+ .unsqueeze(3)
37
+ .permute(1, 0, 2, 3)
38
+ )
39
+
40
+ self.loc.data.copy_(-mean)
41
+ self.scale.data.copy_(1 / (std + 1e-6))
42
+
43
+ def forward(self, input, reverse=False):
44
+ if reverse:
45
+ return self.reverse(input)
46
+ if len(input.shape) == 2:
47
+ input = input[:,:,None,None]
48
+ squeeze = True
49
+ else:
50
+ squeeze = False
51
+
52
+ _, _, height, width = input.shape
53
+
54
+ if self.training and self.initialized.item() == 0:
55
+ self.initialize(input)
56
+ self.initialized.fill_(1)
57
+
58
+ h = self.scale * (input + self.loc)
59
+
60
+ if squeeze:
61
+ h = h.squeeze(-1).squeeze(-1)
62
+
63
+ if self.logdet:
64
+ log_abs = torch.log(torch.abs(self.scale))
65
+ logdet = height*width*torch.sum(log_abs)
66
+ logdet = logdet * torch.ones(input.shape[0]).to(input)
67
+ return h, logdet
68
+
69
+ return h
70
+
71
+ def reverse(self, output):
72
+ if self.training and self.initialized.item() == 0:
73
+ if not self.allow_reverse_init:
74
+ raise RuntimeError(
75
+ "Initializing ActNorm in reverse direction is "
76
+ "disabled by default. Use allow_reverse_init=True to enable."
77
+ )
78
+ else:
79
+ self.initialize(output)
80
+ self.initialized.fill_(1)
81
+
82
+ if len(output.shape) == 2:
83
+ output = output[:,:,None,None]
84
+ squeeze = True
85
+ else:
86
+ squeeze = False
87
+
88
+ h = output / self.scale - self.loc
89
+
90
+ if squeeze:
91
+ h = h.squeeze(-1).squeeze(-1)
92
+ return h
93
+
94
+
95
+ class AbstractEncoder(nn.Module):
96
+ def __init__(self):
97
+ super().__init__()
98
+
99
+ def encode(self, *args, **kwargs):
100
+ raise NotImplementedError
101
+
102
+
103
+ class Labelator(AbstractEncoder):
104
+ """Net2Net Interface for Class-Conditional Model"""
105
+ def __init__(self, n_classes, quantize_interface=True):
106
+ super().__init__()
107
+ self.n_classes = n_classes
108
+ self.quantize_interface = quantize_interface
109
+
110
+ def encode(self, c):
111
+ c = c[:,None]
112
+ if self.quantize_interface:
113
+ return c, None, [None, None, c.long()]
114
+ return c
115
+
116
+
117
+ class SOSProvider(AbstractEncoder):
118
+ # for unconditional training
119
+ def __init__(self, sos_token, quantize_interface=True):
120
+ super().__init__()
121
+ self.sos_token = sos_token
122
+ self.quantize_interface = quantize_interface
123
+
124
+ def encode(self, x):
125
+ # get batch size from data and replicate sos_token
126
+ c = torch.ones(x.shape[0], 1)*self.sos_token
127
+ c = c.long().to(x.device)
128
+ if self.quantize_interface:
129
+ return c, None, [None, None, c]
130
+ return c
ldm/util.py ADDED
@@ -0,0 +1,237 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import importlib
2
+ import os, hashlib
3
+ import requests
4
+
5
+ from tqdm import tqdm
6
+ import torch
7
+ from torch import optim
8
+ import numpy as np
9
+
10
+ from inspect import isfunction
11
+ from PIL import Image, ImageDraw, ImageFont
12
+ URL_MAP = {
13
+ "vgg_lpips": "https://heibox.uni-heidelberg.de/f/607503859c864bc1b30b/?dl=1"
14
+ }
15
+
16
+ CKPT_MAP = {
17
+ "vgg_lpips": "vgg.pth"
18
+ }
19
+
20
+ MD5_MAP = {
21
+ "vgg_lpips": "d507d7349b931f0638a25a48a722f98a"
22
+ }
23
+
24
+ def log_txt_as_img(wh, xc, size=10):
25
+ # wh a tuple of (width, height)
26
+ # xc a list of captions to plot
27
+ b = len(xc)
28
+ txts = list()
29
+ for bi in range(b):
30
+ txt = Image.new("RGB", wh, color="white")
31
+ draw = ImageDraw.Draw(txt)
32
+ font = ImageFont.truetype('font/DejaVuSans.ttf', size=size)
33
+ nc = int(40 * (wh[0] / 256))
34
+ lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc))
35
+
36
+ try:
37
+ draw.text((0, 0), lines, fill="black", font=font)
38
+ except UnicodeEncodeError:
39
+ print("Cant encode string for logging. Skipping.")
40
+
41
+ txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
42
+ txts.append(txt)
43
+ txts = np.stack(txts)
44
+ txts = torch.tensor(txts)
45
+ return txts
46
+
47
+
48
+ def ismap(x):
49
+ if not isinstance(x, torch.Tensor):
50
+ return False
51
+ return (len(x.shape) == 4) and (x.shape[1] > 3)
52
+
53
+
54
+ def isimage(x):
55
+ if not isinstance(x,torch.Tensor):
56
+ return False
57
+ return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
58
+
59
+
60
+ def exists(x):
61
+ return x is not None
62
+
63
+
64
+ def default(val, d):
65
+ if exists(val):
66
+ return val
67
+ return d() if isfunction(d) else d
68
+
69
+ def download(url, local_path, chunk_size=1024):
70
+ os.makedirs(os.path.split(local_path)[0], exist_ok=True)
71
+ with requests.get(url, stream=True) as r:
72
+ total_size = int(r.headers.get("content-length", 0))
73
+ with tqdm(total=total_size, unit="B", unit_scale=True) as pbar:
74
+ with open(local_path, "wb") as f:
75
+ for data in r.iter_content(chunk_size=chunk_size):
76
+ if data:
77
+ f.write(data)
78
+ pbar.update(chunk_size)
79
+
80
+
81
+ def md5_hash(path):
82
+ with open(path, "rb") as f:
83
+ content = f.read()
84
+ return hashlib.md5(content).hexdigest()
85
+
86
+
87
+ def get_ckpt_path(name, root, check=False):
88
+ assert name in URL_MAP
89
+ path = os.path.join(root, CKPT_MAP[name])
90
+ if not os.path.exists(path) or (check and not md5_hash(path) == MD5_MAP[name]):
91
+ print("Downloading {} model from {} to {}".format(name, URL_MAP[name], path))
92
+ download(URL_MAP[name], path)
93
+ md5 = md5_hash(path)
94
+ assert md5 == MD5_MAP[name], md5
95
+ return path
96
+
97
+ def mean_flat(tensor):
98
+ """
99
+ https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
100
+ Take the mean over all non-batch dimensions.
101
+ """
102
+ return tensor.mean(dim=list(range(1, len(tensor.shape))))
103
+
104
+
105
+ def count_params(model, verbose=False):
106
+ total_params = sum(p.numel() for p in model.parameters())
107
+ if verbose:
108
+ print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
109
+ return total_params
110
+
111
+
112
+ def instantiate_from_config(config):
113
+ if not "target" in config:
114
+ if config == '__is_first_stage__':
115
+ return None
116
+ elif config == "__is_unconditional__":
117
+ return None
118
+ raise KeyError("Expected key `target` to instantiate.")
119
+ return get_obj_from_str(config["target"])(**config.get("params", dict()))
120
+
121
+
122
+ def get_obj_from_str(string, reload=False):
123
+ module, cls = string.rsplit(".", 1)
124
+ if reload:
125
+ module_imp = importlib.import_module(module)
126
+ importlib.reload(module_imp)
127
+ return getattr(importlib.import_module(module, package=None), cls)
128
+
129
+
130
+ class AdamWwithEMAandWings(optim.Optimizer):
131
+ # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298
132
+ def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8, # TODO: check hyperparameters before using
133
+ weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999, # ema decay to match previous code
134
+ ema_power=1., param_names=()):
135
+ """AdamW that saves EMA versions of the parameters."""
136
+ if not 0.0 <= lr:
137
+ raise ValueError("Invalid learning rate: {}".format(lr))
138
+ if not 0.0 <= eps:
139
+ raise ValueError("Invalid epsilon value: {}".format(eps))
140
+ if not 0.0 <= betas[0] < 1.0:
141
+ raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
142
+ if not 0.0 <= betas[1] < 1.0:
143
+ raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
144
+ if not 0.0 <= weight_decay:
145
+ raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
146
+ if not 0.0 <= ema_decay <= 1.0:
147
+ raise ValueError("Invalid ema_decay value: {}".format(ema_decay))
148
+ defaults = dict(lr=lr, betas=betas, eps=eps,
149
+ weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay,
150
+ ema_power=ema_power, param_names=param_names)
151
+ super().__init__(params, defaults)
152
+
153
+ def __setstate__(self, state):
154
+ super().__setstate__(state)
155
+ for group in self.param_groups:
156
+ group.setdefault('amsgrad', False)
157
+
158
+ @torch.no_grad()
159
+ def step(self, closure=None):
160
+ """Performs a single optimization step.
161
+ Args:
162
+ closure (callable, optional): A closure that reevaluates the model
163
+ and returns the loss.
164
+ """
165
+ loss = None
166
+ if closure is not None:
167
+ with torch.enable_grad():
168
+ loss = closure()
169
+
170
+ for group in self.param_groups:
171
+ params_with_grad = []
172
+ grads = []
173
+ exp_avgs = []
174
+ exp_avg_sqs = []
175
+ ema_params_with_grad = []
176
+ state_sums = []
177
+ max_exp_avg_sqs = []
178
+ state_steps = []
179
+ amsgrad = group['amsgrad']
180
+ beta1, beta2 = group['betas']
181
+ ema_decay = group['ema_decay']
182
+ ema_power = group['ema_power']
183
+
184
+ for p in group['params']:
185
+ if p.grad is None:
186
+ continue
187
+ params_with_grad.append(p)
188
+ if p.grad.is_sparse:
189
+ raise RuntimeError('AdamW does not support sparse gradients')
190
+ grads.append(p.grad)
191
+
192
+ state = self.state[p]
193
+
194
+ # State initialization
195
+ if len(state) == 0:
196
+ state['step'] = 0
197
+ # Exponential moving average of gradient values
198
+ state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
199
+ # Exponential moving average of squared gradient values
200
+ state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
201
+ if amsgrad:
202
+ # Maintains max of all exp. moving avg. of sq. grad. values
203
+ state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
204
+ # Exponential moving average of parameter values
205
+ state['param_exp_avg'] = p.detach().float().clone()
206
+
207
+ exp_avgs.append(state['exp_avg'])
208
+ exp_avg_sqs.append(state['exp_avg_sq'])
209
+ ema_params_with_grad.append(state['param_exp_avg'])
210
+
211
+ if amsgrad:
212
+ max_exp_avg_sqs.append(state['max_exp_avg_sq'])
213
+
214
+ # update the steps for each param group update
215
+ state['step'] += 1
216
+ # record the step after step update
217
+ state_steps.append(state['step'])
218
+
219
+ optim._functional.adamw(params_with_grad,
220
+ grads,
221
+ exp_avgs,
222
+ exp_avg_sqs,
223
+ max_exp_avg_sqs,
224
+ state_steps,
225
+ amsgrad=amsgrad,
226
+ beta1=beta1,
227
+ beta2=beta2,
228
+ lr=group['lr'],
229
+ weight_decay=group['weight_decay'],
230
+ eps=group['eps'],
231
+ maximize=False)
232
+
233
+ cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power)
234
+ for param, ema_param in zip(params_with_grad, ema_params_with_grad):
235
+ ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay)
236
+
237
+ return loss