File size: 10,033 Bytes
af7c068 |
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 |
# coding=utf-8
# Copyright 2022 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 inspect
import tempfile
import unittest
from typing import Dict, List, Tuple
import numpy as np
import torch
from diffusers.modeling_utils import ModelMixin
from diffusers.training_utils import EMAModel
from diffusers.utils import torch_device
class ModelTesterMixin:
def test_from_pretrained_save_pretrained(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
new_model = self.model_class.from_pretrained(tmpdirname)
new_model.to(torch_device)
with torch.no_grad():
# Warmup pass when using mps (see #372)
if torch_device == "mps" and isinstance(model, ModelMixin):
_ = model(**self.dummy_input)
_ = new_model(**self.dummy_input)
image = model(**inputs_dict)
if isinstance(image, dict):
image = image.sample
new_image = new_model(**inputs_dict)
if isinstance(new_image, dict):
new_image = new_image.sample
max_diff = (image - new_image).abs().sum().item()
self.assertLessEqual(max_diff, 5e-5, "Models give different forward passes")
def test_determinism(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()
with torch.no_grad():
# Warmup pass when using mps (see #372)
if torch_device == "mps" and isinstance(model, ModelMixin):
model(**self.dummy_input)
first = model(**inputs_dict)
if isinstance(first, dict):
first = first.sample
second = model(**inputs_dict)
if isinstance(second, dict):
second = second.sample
out_1 = first.cpu().numpy()
out_2 = second.cpu().numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_output(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()
with torch.no_grad():
output = model(**inputs_dict)
if isinstance(output, dict):
output = output.sample
self.assertIsNotNone(output)
expected_shape = inputs_dict["sample"].shape
self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
def test_forward_with_norm_groups(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
init_dict["norm_num_groups"] = 16
init_dict["block_out_channels"] = (16, 32)
model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()
with torch.no_grad():
output = model(**inputs_dict)
if isinstance(output, dict):
output = output.sample
self.assertIsNotNone(output)
expected_shape = inputs_dict["sample"].shape
self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
def test_forward_signature(self):
init_dict, _ = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = ["sample", "timestep"]
self.assertListEqual(arg_names[:2], expected_arg_names)
def test_model_from_pretrained(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()
# test if the model can be loaded from the config
# and has all the expected shape
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
new_model = self.model_class.from_pretrained(tmpdirname)
new_model.to(torch_device)
new_model.eval()
# check if all parameters shape are the same
for param_name in model.state_dict().keys():
param_1 = model.state_dict()[param_name]
param_2 = new_model.state_dict()[param_name]
self.assertEqual(param_1.shape, param_2.shape)
with torch.no_grad():
output_1 = model(**inputs_dict)
if isinstance(output_1, dict):
output_1 = output_1.sample
output_2 = new_model(**inputs_dict)
if isinstance(output_2, dict):
output_2 = output_2.sample
self.assertEqual(output_1.shape, output_2.shape)
@unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
def test_training(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.train()
output = model(**inputs_dict)
if isinstance(output, dict):
output = output.sample
noise = torch.randn((inputs_dict["sample"].shape[0],) + self.output_shape).to(torch_device)
loss = torch.nn.functional.mse_loss(output, noise)
loss.backward()
@unittest.skipIf(torch_device == "mps", "Training is not supported in mps")
def test_ema_training(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.train()
ema_model = EMAModel(model, device=torch_device)
output = model(**inputs_dict)
if isinstance(output, dict):
output = output.sample
noise = torch.randn((inputs_dict["sample"].shape[0],) + self.output_shape).to(torch_device)
loss = torch.nn.functional.mse_loss(output, noise)
loss.backward()
ema_model.step(model)
def test_outputs_equivalence(self):
def set_nan_tensor_to_zero(t):
# Temporary fallback until `aten::_index_put_impl_` is implemented in mps
# Track progress in https://github.com/pytorch/pytorch/issues/77764
device = t.device
if device.type == "mps":
t = t.to("cpu")
t[t != t] = 0
return t.to(device)
def recursive_check(tuple_object, dict_object):
if isinstance(tuple_object, (List, Tuple)):
for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object.values()):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif isinstance(tuple_object, Dict):
for tuple_iterable_value, dict_iterable_value in zip(tuple_object.values(), dict_object.values()):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif tuple_object is None:
return
else:
self.assertTrue(
torch.allclose(
set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
),
msg=(
"Tuple and dict output are not equal. Difference:"
f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
),
)
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()
with torch.no_grad():
# Warmup pass when using mps (see #372)
if torch_device == "mps" and isinstance(model, ModelMixin):
model(**self.dummy_input)
outputs_dict = model(**inputs_dict)
outputs_tuple = model(**inputs_dict, return_dict=False)
recursive_check(outputs_tuple, outputs_dict)
@unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")
def test_enable_disable_gradient_checkpointing(self):
if not self.model_class._supports_gradient_checkpointing:
return # Skip test if model does not support gradient checkpointing
init_dict, _ = self.prepare_init_args_and_inputs_for_common()
# at init model should have gradient checkpointing disabled
model = self.model_class(**init_dict)
self.assertFalse(model.is_gradient_checkpointing)
# check enable works
model.enable_gradient_checkpointing()
self.assertTrue(model.is_gradient_checkpointing)
# check disable works
model.disable_gradient_checkpointing()
self.assertFalse(model.is_gradient_checkpointing)
|