File size: 5,932 Bytes
b334e29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
'''
 * Copyright (c) 2023 Salesforce, Inc.
 * All rights reserved.
 * SPDX-License-Identifier: Apache License 2.0
 * For full license text, see LICENSE.txt file in the repo root or http://www.apache.org/licenses/
 * By Can Qin
 * Modified from ControlNet repo: https://github.com/lllyasviel/ControlNet
 * Copyright (c) 2023 Lvmin Zhang and Maneesh Agrawala
'''

# based on https://github.com/isl-org/MiDaS

import cv2
import os
import torch
import torch.nn as nn
from torchvision.transforms import Compose

from .midas.dpt_depth import DPTDepthModel
from .midas.midas_net import MidasNet
from .midas.midas_net_custom import MidasNet_small
from .midas.transforms import Resize, NormalizeImage, PrepareForNet
from annotator.util import annotator_ckpts_path


ISL_PATHS = {
    "dpt_large": os.path.join(annotator_ckpts_path, "dpt_large_384.pt"),
    "dpt_hybrid": os.path.join(annotator_ckpts_path, "dpt_hybrid-midas-501f0c75.pt"),
    "midas_v21": "",
    "midas_v21_small": "",
}

# remote_model_path = "https://huggingface.co/lllyasviel/ControlNet/resolve/main/annotator/ckpts/dpt_hybrid-midas-501f0c75.pt"
remote_model_path = "https://huggingface.co/Salesforce/UniControl/blob/main/annotator/ckpts/dpt_large_384.pt"

def disabled_train(self, mode=True):
    """Overwrite model.train with this function to make sure train/eval mode
    does not change anymore."""
    return self


def load_midas_transform(model_type):
    # https://github.com/isl-org/MiDaS/blob/master/run.py
    # load transform only
    if model_type == "dpt_large":  # DPT-Large
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "dpt_hybrid":  # DPT-Hybrid
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "midas_v21":
        net_w, net_h = 384, 384
        resize_mode = "upper_bound"
        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

    elif model_type == "midas_v21_small":
        net_w, net_h = 256, 256
        resize_mode = "upper_bound"
        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

    else:
        assert False, f"model_type '{model_type}' not implemented, use: --model_type large"

    transform = Compose(
        [
            Resize(
                net_w,
                net_h,
                resize_target=None,
                keep_aspect_ratio=True,
                ensure_multiple_of=32,
                resize_method=resize_mode,
                image_interpolation_method=cv2.INTER_CUBIC,
            ),
            normalization,
            PrepareForNet(),
        ]
    )

    return transform


def load_model(model_type):
    # https://github.com/isl-org/MiDaS/blob/master/run.py
    # load network
    model_path = ISL_PATHS[model_type]
    if model_type == "dpt_large":  # DPT-Large
        # if not os.path.exists(model_path):
        #     from basicsr.utils.download_util import load_file_from_url
        #     load_file_from_url(remote_model_path, model_dir=annotator_ckpts_path)
        model_path = remote_model_path
        model = DPTDepthModel(
            path=model_path,
            backbone="vitl16_384",
            non_negative=True,
        )
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "dpt_hybrid":  # DPT-Hybrid
        if not os.path.exists(model_path):
            from basicsr.utils.download_util import load_file_from_url
            load_file_from_url(remote_model_path, model_dir=annotator_ckpts_path)

        model = DPTDepthModel(
            path=model_path,
            backbone="vitb_rn50_384",
            non_negative=True,
        )
        net_w, net_h = 384, 384
        resize_mode = "minimal"
        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

    elif model_type == "midas_v21":
        model = MidasNet(model_path, non_negative=True)
        net_w, net_h = 384, 384
        resize_mode = "upper_bound"
        normalization = NormalizeImage(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
        )

    elif model_type == "midas_v21_small":
        model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
                               non_negative=True, blocks={'expand': True})
        net_w, net_h = 256, 256
        resize_mode = "upper_bound"
        normalization = NormalizeImage(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
        )

    else:
        print(f"model_type '{model_type}' not implemented, use: --model_type large")
        assert False

    transform = Compose(
        [
            Resize(
                net_w,
                net_h,
                resize_target=None,
                keep_aspect_ratio=True,
                ensure_multiple_of=32,
                resize_method=resize_mode,
                image_interpolation_method=cv2.INTER_CUBIC,
            ),
            normalization,
            PrepareForNet(),
        ]
    )

    return model.eval(), transform


class MiDaSInference(nn.Module):
    MODEL_TYPES_TORCH_HUB = [
        "DPT_Large",
        "DPT_Hybrid",
        "MiDaS_small"
    ]
    MODEL_TYPES_ISL = [
        "dpt_large",
        "dpt_hybrid",
        "midas_v21",
        "midas_v21_small",
    ]

    def __init__(self, model_type):
        super().__init__()
        assert (model_type in self.MODEL_TYPES_ISL)
        model, _ = load_model(model_type)
        self.model = model
        self.model.train = disabled_train

    def forward(self, x):
        with torch.no_grad():
            prediction = self.model(x)
        return prediction