File size: 16,104 Bytes
43b7e92 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 |
import tempfile
from io import BytesIO
import requests
import torch
from huggingface_hub import hf_hub_download, snapshot_download
from diffusers.models.attention_processor import AttnProcessor
from diffusers.utils.testing_utils import (
numpy_cosine_similarity_distance,
torch_device,
)
def download_single_file_checkpoint(repo_id, filename, tmpdir):
path = hf_hub_download(repo_id, filename=filename, local_dir=tmpdir)
return path
def download_original_config(config_url, tmpdir):
original_config_file = BytesIO(requests.get(config_url).content)
path = f"{tmpdir}/config.yaml"
with open(path, "wb") as f:
f.write(original_config_file.read())
return path
def download_diffusers_config(repo_id, tmpdir):
path = snapshot_download(
repo_id,
ignore_patterns=[
"**/*.ckpt",
"*.ckpt",
"**/*.bin",
"*.bin",
"**/*.pt",
"*.pt",
"**/*.safetensors",
"*.safetensors",
],
allow_patterns=["**/*.json", "*.json", "*.txt", "**/*.txt"],
local_dir=tmpdir,
)
return path
class SDSingleFileTesterMixin:
def _compare_component_configs(self, pipe, single_file_pipe):
for param_name, param_value in single_file_pipe.text_encoder.config.to_dict().items():
if param_name in ["torch_dtype", "architectures", "_name_or_path"]:
continue
assert pipe.text_encoder.config.to_dict()[param_name] == param_value
PARAMS_TO_IGNORE = [
"torch_dtype",
"_name_or_path",
"architectures",
"_use_default_values",
"_diffusers_version",
]
for component_name, component in single_file_pipe.components.items():
if component_name in single_file_pipe._optional_components:
continue
# skip testing transformer based components here
# skip text encoders / safety checkers since they have already been tested
if component_name in ["text_encoder", "tokenizer", "safety_checker", "feature_extractor"]:
continue
assert component_name in pipe.components, f"single file {component_name} not found in pretrained pipeline"
assert isinstance(
component, pipe.components[component_name].__class__
), f"single file {component.__class__.__name__} and pretrained {pipe.components[component_name].__class__.__name__} are not the same"
for param_name, param_value in component.config.items():
if param_name in PARAMS_TO_IGNORE:
continue
# Some pretrained configs will set upcast attention to None
# In single file loading it defaults to the value in the class __init__ which is False
if param_name == "upcast_attention" and pipe.components[component_name].config[param_name] is None:
pipe.components[component_name].config[param_name] = param_value
assert (
pipe.components[component_name].config[param_name] == param_value
), f"single file {param_name}: {param_value} differs from pretrained {pipe.components[component_name].config[param_name]}"
def test_single_file_components(self, pipe=None, single_file_pipe=None):
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
self.ckpt_path, safety_checker=None
)
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_components_local_files_only(self, pipe=None, single_file_pipe=None):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
with tempfile.TemporaryDirectory() as tmpdir:
ckpt_filename = self.ckpt_path.split("/")[-1]
local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
local_ckpt_path, safety_checker=None, local_files_only=True
)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_components_with_original_config(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
# Not possible to infer this value when original config is provided
# we just pass it in here otherwise this test will fail
upcast_attention = pipe.unet.config.upcast_attention
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
self.ckpt_path,
original_config=self.original_config,
safety_checker=None,
upcast_attention=upcast_attention,
)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_components_with_original_config_local_files_only(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
# Not possible to infer this value when original config is provided
# we just pass it in here otherwise this test will fail
upcast_attention = pipe.unet.config.upcast_attention
with tempfile.TemporaryDirectory() as tmpdir:
ckpt_filename = self.ckpt_path.split("/")[-1]
local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
local_original_config = download_original_config(self.original_config, tmpdir)
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
local_ckpt_path,
original_config=local_original_config,
safety_checker=None,
upcast_attention=upcast_attention,
local_files_only=True,
)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_format_inference_is_same_as_pretrained(self, expected_max_diff=1e-4):
sf_pipe = self.pipeline_class.from_single_file(self.ckpt_path, safety_checker=None)
sf_pipe.unet.set_attn_processor(AttnProcessor())
sf_pipe.enable_model_cpu_offload()
inputs = self.get_inputs(torch_device)
image_single_file = sf_pipe(**inputs).images[0]
pipe = self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
pipe.unet.set_attn_processor(AttnProcessor())
pipe.enable_model_cpu_offload()
inputs = self.get_inputs(torch_device)
image = pipe(**inputs).images[0]
max_diff = numpy_cosine_similarity_distance(image.flatten(), image_single_file.flatten())
assert max_diff < expected_max_diff
def test_single_file_components_with_diffusers_config(
self,
pipe=None,
single_file_pipe=None,
):
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
self.ckpt_path, config=self.repo_id, safety_checker=None
)
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_components_with_diffusers_config_local_files_only(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
with tempfile.TemporaryDirectory() as tmpdir:
ckpt_filename = self.ckpt_path.split("/")[-1]
local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
local_diffusers_config = download_diffusers_config(self.repo_id, tmpdir)
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
local_ckpt_path, config=local_diffusers_config, safety_checker=None, local_files_only=True
)
self._compare_component_configs(pipe, single_file_pipe)
class SDXLSingleFileTesterMixin:
def _compare_component_configs(self, pipe, single_file_pipe):
# Skip testing the text_encoder for Refiner Pipelines
if pipe.text_encoder:
for param_name, param_value in single_file_pipe.text_encoder.config.to_dict().items():
if param_name in ["torch_dtype", "architectures", "_name_or_path"]:
continue
assert pipe.text_encoder.config.to_dict()[param_name] == param_value
for param_name, param_value in single_file_pipe.text_encoder_2.config.to_dict().items():
if param_name in ["torch_dtype", "architectures", "_name_or_path"]:
continue
assert pipe.text_encoder_2.config.to_dict()[param_name] == param_value
PARAMS_TO_IGNORE = [
"torch_dtype",
"_name_or_path",
"architectures",
"_use_default_values",
"_diffusers_version",
]
for component_name, component in single_file_pipe.components.items():
if component_name in single_file_pipe._optional_components:
continue
# skip text encoders since they have already been tested
if component_name in ["text_encoder", "text_encoder_2", "tokenizer", "tokenizer_2"]:
continue
# skip safety checker if it is not present in the pipeline
if component_name in ["safety_checker", "feature_extractor"]:
continue
assert component_name in pipe.components, f"single file {component_name} not found in pretrained pipeline"
assert isinstance(
component, pipe.components[component_name].__class__
), f"single file {component.__class__.__name__} and pretrained {pipe.components[component_name].__class__.__name__} are not the same"
for param_name, param_value in component.config.items():
if param_name in PARAMS_TO_IGNORE:
continue
# Some pretrained configs will set upcast attention to None
# In single file loading it defaults to the value in the class __init__ which is False
if param_name == "upcast_attention" and pipe.components[component_name].config[param_name] is None:
pipe.components[component_name].config[param_name] = param_value
assert (
pipe.components[component_name].config[param_name] == param_value
), f"single file {param_name}: {param_value} differs from pretrained {pipe.components[component_name].config[param_name]}"
def test_single_file_components(self, pipe=None, single_file_pipe=None):
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
self.ckpt_path, safety_checker=None
)
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
self._compare_component_configs(
pipe,
single_file_pipe,
)
def test_single_file_components_local_files_only(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
with tempfile.TemporaryDirectory() as tmpdir:
ckpt_filename = self.ckpt_path.split("/")[-1]
local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
local_ckpt_path, safety_checker=None, local_files_only=True
)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_components_with_original_config(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
# Not possible to infer this value when original config is provided
# we just pass it in here otherwise this test will fail
upcast_attention = pipe.unet.config.upcast_attention
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
self.ckpt_path,
original_config=self.original_config,
safety_checker=None,
upcast_attention=upcast_attention,
)
self._compare_component_configs(
pipe,
single_file_pipe,
)
def test_single_file_components_with_original_config_local_files_only(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
# Not possible to infer this value when original config is provided
# we just pass it in here otherwise this test will fail
upcast_attention = pipe.unet.config.upcast_attention
with tempfile.TemporaryDirectory() as tmpdir:
ckpt_filename = self.ckpt_path.split("/")[-1]
local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
local_original_config = download_original_config(self.original_config, tmpdir)
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
local_ckpt_path,
original_config=local_original_config,
upcast_attention=upcast_attention,
safety_checker=None,
local_files_only=True,
)
self._compare_component_configs(
pipe,
single_file_pipe,
)
def test_single_file_components_with_diffusers_config(
self,
pipe=None,
single_file_pipe=None,
):
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
self.ckpt_path, config=self.repo_id, safety_checker=None
)
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_components_with_diffusers_config_local_files_only(
self,
pipe=None,
single_file_pipe=None,
):
pipe = pipe or self.pipeline_class.from_pretrained(self.repo_id, safety_checker=None)
with tempfile.TemporaryDirectory() as tmpdir:
ckpt_filename = self.ckpt_path.split("/")[-1]
local_ckpt_path = download_single_file_checkpoint(self.repo_id, ckpt_filename, tmpdir)
local_diffusers_config = download_diffusers_config(self.repo_id, tmpdir)
single_file_pipe = single_file_pipe or self.pipeline_class.from_single_file(
local_ckpt_path, config=local_diffusers_config, safety_checker=None, local_files_only=True
)
self._compare_component_configs(pipe, single_file_pipe)
def test_single_file_format_inference_is_same_as_pretrained(self, expected_max_diff=1e-4):
sf_pipe = self.pipeline_class.from_single_file(self.ckpt_path, torch_dtype=torch.float16, safety_checker=None)
sf_pipe.unet.set_default_attn_processor()
sf_pipe.enable_model_cpu_offload()
inputs = self.get_inputs(torch_device)
image_single_file = sf_pipe(**inputs).images[0]
pipe = self.pipeline_class.from_pretrained(self.repo_id, torch_dtype=torch.float16, safety_checker=None)
pipe.unet.set_default_attn_processor()
pipe.enable_model_cpu_offload()
inputs = self.get_inputs(torch_device)
image = pipe(**inputs).images[0]
max_diff = numpy_cosine_similarity_distance(image.flatten(), image_single_file.flatten())
assert max_diff < expected_max_diff
|