File size: 12,851 Bytes
3a25a0a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 |
# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import tempfile
import unittest
import torch
from diffusers import UNet2DConditionModel
from diffusers.training_utils import EMAModel
from diffusers.utils.testing_utils import enable_full_determinism, skip_mps, torch_device
enable_full_determinism()
class EMAModelTests(unittest.TestCase):
model_id = "hf-internal-testing/tiny-stable-diffusion-pipe"
batch_size = 1
prompt_length = 77
text_encoder_hidden_dim = 32
num_in_channels = 4
latent_height = latent_width = 64
generator = torch.manual_seed(0)
def get_models(self, decay=0.9999):
unet = UNet2DConditionModel.from_pretrained(self.model_id, subfolder="unet")
unet = unet.to(torch_device)
ema_unet = EMAModel(unet.parameters(), decay=decay, model_cls=UNet2DConditionModel, model_config=unet.config)
return unet, ema_unet
def get_dummy_inputs(self):
noisy_latents = torch.randn(
self.batch_size, self.num_in_channels, self.latent_height, self.latent_width, generator=self.generator
).to(torch_device)
timesteps = torch.randint(0, 1000, size=(self.batch_size,), generator=self.generator).to(torch_device)
encoder_hidden_states = torch.randn(
self.batch_size, self.prompt_length, self.text_encoder_hidden_dim, generator=self.generator
).to(torch_device)
return noisy_latents, timesteps, encoder_hidden_states
def simulate_backprop(self, unet):
updated_state_dict = {}
for k, param in unet.state_dict().items():
updated_param = torch.randn_like(param) + (param * torch.randn_like(param))
updated_state_dict.update({k: updated_param})
unet.load_state_dict(updated_state_dict)
return unet
def test_optimization_steps_updated(self):
unet, ema_unet = self.get_models()
# Take the first (hypothetical) EMA step.
ema_unet.step(unet.parameters())
assert ema_unet.optimization_step == 1
# Take two more.
for _ in range(2):
ema_unet.step(unet.parameters())
assert ema_unet.optimization_step == 3
def test_shadow_params_not_updated(self):
unet, ema_unet = self.get_models()
# Since the `unet` is not being updated (i.e., backprop'd)
# there won't be any difference between the `params` of `unet`
# and `ema_unet` even if we call `ema_unet.step(unet.parameters())`.
ema_unet.step(unet.parameters())
orig_params = list(unet.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert torch.allclose(s_param, param)
# The above holds true even if we call `ema.step()` multiple times since
# `unet` params are still not being updated.
for _ in range(4):
ema_unet.step(unet.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert torch.allclose(s_param, param)
def test_shadow_params_updated(self):
unet, ema_unet = self.get_models()
# Here we simulate the parameter updates for `unet`. Since there might
# be some parameters which are initialized to zero we take extra care to
# initialize their values to something non-zero before the multiplication.
unet_pseudo_updated_step_one = self.simulate_backprop(unet)
# Take the EMA step.
ema_unet.step(unet_pseudo_updated_step_one.parameters())
# Now the EMA'd parameters won't be equal to the original model parameters.
orig_params = list(unet_pseudo_updated_step_one.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert ~torch.allclose(s_param, param)
# Ensure this is the case when we take multiple EMA steps.
for _ in range(4):
ema_unet.step(unet.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert ~torch.allclose(s_param, param)
def test_consecutive_shadow_params_updated(self):
# If we call EMA step after a backpropagation consecutively for two times,
# the shadow params from those two steps should be different.
unet, ema_unet = self.get_models()
# First backprop + EMA
unet_step_one = self.simulate_backprop(unet)
ema_unet.step(unet_step_one.parameters())
step_one_shadow_params = ema_unet.shadow_params
# Second backprop + EMA
unet_step_two = self.simulate_backprop(unet_step_one)
ema_unet.step(unet_step_two.parameters())
step_two_shadow_params = ema_unet.shadow_params
for step_one, step_two in zip(step_one_shadow_params, step_two_shadow_params):
assert ~torch.allclose(step_one, step_two)
def test_zero_decay(self):
# If there's no decay even if there are backprops, EMA steps
# won't take any effect i.e., the shadow params would remain the
# same.
unet, ema_unet = self.get_models(decay=0.0)
unet_step_one = self.simulate_backprop(unet)
ema_unet.step(unet_step_one.parameters())
step_one_shadow_params = ema_unet.shadow_params
unet_step_two = self.simulate_backprop(unet_step_one)
ema_unet.step(unet_step_two.parameters())
step_two_shadow_params = ema_unet.shadow_params
for step_one, step_two in zip(step_one_shadow_params, step_two_shadow_params):
assert torch.allclose(step_one, step_two)
@skip_mps
def test_serialization(self):
unet, ema_unet = self.get_models()
noisy_latents, timesteps, encoder_hidden_states = self.get_dummy_inputs()
with tempfile.TemporaryDirectory() as tmpdir:
ema_unet.save_pretrained(tmpdir)
loaded_unet = UNet2DConditionModel.from_pretrained(tmpdir, model_cls=UNet2DConditionModel)
loaded_unet = loaded_unet.to(unet.device)
# Since no EMA step has been performed the outputs should match.
output = unet(noisy_latents, timesteps, encoder_hidden_states).sample
output_loaded = loaded_unet(noisy_latents, timesteps, encoder_hidden_states).sample
assert torch.allclose(output, output_loaded, atol=1e-4)
class EMAModelTestsForeach(unittest.TestCase):
model_id = "hf-internal-testing/tiny-stable-diffusion-pipe"
batch_size = 1
prompt_length = 77
text_encoder_hidden_dim = 32
num_in_channels = 4
latent_height = latent_width = 64
generator = torch.manual_seed(0)
def get_models(self, decay=0.9999):
unet = UNet2DConditionModel.from_pretrained(self.model_id, subfolder="unet")
unet = unet.to(torch_device)
ema_unet = EMAModel(
unet.parameters(), decay=decay, model_cls=UNet2DConditionModel, model_config=unet.config, foreach=True
)
return unet, ema_unet
def get_dummy_inputs(self):
noisy_latents = torch.randn(
self.batch_size, self.num_in_channels, self.latent_height, self.latent_width, generator=self.generator
).to(torch_device)
timesteps = torch.randint(0, 1000, size=(self.batch_size,), generator=self.generator).to(torch_device)
encoder_hidden_states = torch.randn(
self.batch_size, self.prompt_length, self.text_encoder_hidden_dim, generator=self.generator
).to(torch_device)
return noisy_latents, timesteps, encoder_hidden_states
def simulate_backprop(self, unet):
updated_state_dict = {}
for k, param in unet.state_dict().items():
updated_param = torch.randn_like(param) + (param * torch.randn_like(param))
updated_state_dict.update({k: updated_param})
unet.load_state_dict(updated_state_dict)
return unet
def test_optimization_steps_updated(self):
unet, ema_unet = self.get_models()
# Take the first (hypothetical) EMA step.
ema_unet.step(unet.parameters())
assert ema_unet.optimization_step == 1
# Take two more.
for _ in range(2):
ema_unet.step(unet.parameters())
assert ema_unet.optimization_step == 3
def test_shadow_params_not_updated(self):
unet, ema_unet = self.get_models()
# Since the `unet` is not being updated (i.e., backprop'd)
# there won't be any difference between the `params` of `unet`
# and `ema_unet` even if we call `ema_unet.step(unet.parameters())`.
ema_unet.step(unet.parameters())
orig_params = list(unet.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert torch.allclose(s_param, param)
# The above holds true even if we call `ema.step()` multiple times since
# `unet` params are still not being updated.
for _ in range(4):
ema_unet.step(unet.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert torch.allclose(s_param, param)
def test_shadow_params_updated(self):
unet, ema_unet = self.get_models()
# Here we simulate the parameter updates for `unet`. Since there might
# be some parameters which are initialized to zero we take extra care to
# initialize their values to something non-zero before the multiplication.
unet_pseudo_updated_step_one = self.simulate_backprop(unet)
# Take the EMA step.
ema_unet.step(unet_pseudo_updated_step_one.parameters())
# Now the EMA'd parameters won't be equal to the original model parameters.
orig_params = list(unet_pseudo_updated_step_one.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert ~torch.allclose(s_param, param)
# Ensure this is the case when we take multiple EMA steps.
for _ in range(4):
ema_unet.step(unet.parameters())
for s_param, param in zip(ema_unet.shadow_params, orig_params):
assert ~torch.allclose(s_param, param)
def test_consecutive_shadow_params_updated(self):
# If we call EMA step after a backpropagation consecutively for two times,
# the shadow params from those two steps should be different.
unet, ema_unet = self.get_models()
# First backprop + EMA
unet_step_one = self.simulate_backprop(unet)
ema_unet.step(unet_step_one.parameters())
step_one_shadow_params = ema_unet.shadow_params
# Second backprop + EMA
unet_step_two = self.simulate_backprop(unet_step_one)
ema_unet.step(unet_step_two.parameters())
step_two_shadow_params = ema_unet.shadow_params
for step_one, step_two in zip(step_one_shadow_params, step_two_shadow_params):
assert ~torch.allclose(step_one, step_two)
def test_zero_decay(self):
# If there's no decay even if there are backprops, EMA steps
# won't take any effect i.e., the shadow params would remain the
# same.
unet, ema_unet = self.get_models(decay=0.0)
unet_step_one = self.simulate_backprop(unet)
ema_unet.step(unet_step_one.parameters())
step_one_shadow_params = ema_unet.shadow_params
unet_step_two = self.simulate_backprop(unet_step_one)
ema_unet.step(unet_step_two.parameters())
step_two_shadow_params = ema_unet.shadow_params
for step_one, step_two in zip(step_one_shadow_params, step_two_shadow_params):
assert torch.allclose(step_one, step_two)
@skip_mps
def test_serialization(self):
unet, ema_unet = self.get_models()
noisy_latents, timesteps, encoder_hidden_states = self.get_dummy_inputs()
with tempfile.TemporaryDirectory() as tmpdir:
ema_unet.save_pretrained(tmpdir)
loaded_unet = UNet2DConditionModel.from_pretrained(tmpdir, model_cls=UNet2DConditionModel)
loaded_unet = loaded_unet.to(unet.device)
# Since no EMA step has been performed the outputs should match.
output = unet(noisy_latents, timesteps, encoder_hidden_states).sample
output_loaded = loaded_unet(noisy_latents, timesteps, encoder_hidden_states).sample
assert torch.allclose(output, output_loaded, atol=1e-4)
|