File size: 17,409 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 |
# 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 gc
import random
import tempfile
import unittest
import numpy as np
import torch
from PIL import Image
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, DDIMScheduler, DDPMScheduler, StableDiffusionUpscalePipeline, UNet2DConditionModel
from diffusers.utils.testing_utils import (
enable_full_determinism,
floats_tensor,
load_image,
load_numpy,
require_torch_gpu,
slow,
torch_device,
)
enable_full_determinism()
class StableDiffusionUpscalePipelineFastTests(unittest.TestCase):
def setUp(self):
# clean up the VRAM before each test
super().setUp()
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
@property
def dummy_image(self):
batch_size = 1
num_channels = 3
sizes = (32, 32)
image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device)
return image
@property
def dummy_cond_unet_upscale(self):
torch.manual_seed(0)
model = UNet2DConditionModel(
block_out_channels=(32, 32, 64),
layers_per_block=2,
sample_size=32,
in_channels=7,
out_channels=4,
down_block_types=("DownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D"),
up_block_types=("CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"),
cross_attention_dim=32,
# SD2-specific config below
attention_head_dim=8,
use_linear_projection=True,
only_cross_attention=(True, True, False),
num_class_embeds=100,
)
return model
@property
def dummy_vae(self):
torch.manual_seed(0)
model = AutoencoderKL(
block_out_channels=[32, 32, 64],
in_channels=3,
out_channels=3,
down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D", "DownEncoderBlock2D"],
up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D", "UpDecoderBlock2D"],
latent_channels=4,
)
return model
@property
def dummy_text_encoder(self):
torch.manual_seed(0)
config = CLIPTextConfig(
bos_token_id=0,
eos_token_id=2,
hidden_size=32,
intermediate_size=37,
layer_norm_eps=1e-05,
num_attention_heads=4,
num_hidden_layers=5,
pad_token_id=1,
vocab_size=1000,
# SD2-specific config below
hidden_act="gelu",
projection_dim=512,
)
return CLIPTextModel(config)
def test_stable_diffusion_upscale(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
unet = self.dummy_cond_unet_upscale
low_res_scheduler = DDPMScheduler()
scheduler = DDIMScheduler(prediction_type="v_prediction")
vae = self.dummy_vae
text_encoder = self.dummy_text_encoder
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
# make sure here that pndm scheduler skips prk
sd_pipe = StableDiffusionUpscalePipeline(
unet=unet,
low_res_scheduler=low_res_scheduler,
scheduler=scheduler,
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
max_noise_level=350,
)
sd_pipe = sd_pipe.to(device)
sd_pipe.set_progress_bar_config(disable=None)
prompt = "A painting of a squirrel eating a burger"
generator = torch.Generator(device=device).manual_seed(0)
output = sd_pipe(
[prompt],
image=low_res_image,
generator=generator,
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
)
image = output.images
generator = torch.Generator(device=device).manual_seed(0)
image_from_tuple = sd_pipe(
[prompt],
image=low_res_image,
generator=generator,
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
return_dict=False,
)[0]
image_slice = image[0, -3:, -3:, -1]
image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
expected_height_width = low_res_image.size[0] * 4
assert image.shape == (1, expected_height_width, expected_height_width, 3)
expected_slice = np.array([0.3113, 0.3910, 0.4272, 0.4859, 0.5061, 0.4652, 0.5362, 0.5715, 0.5661])
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2
def test_stable_diffusion_upscale_batch(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
unet = self.dummy_cond_unet_upscale
low_res_scheduler = DDPMScheduler()
scheduler = DDIMScheduler(prediction_type="v_prediction")
vae = self.dummy_vae
text_encoder = self.dummy_text_encoder
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
# make sure here that pndm scheduler skips prk
sd_pipe = StableDiffusionUpscalePipeline(
unet=unet,
low_res_scheduler=low_res_scheduler,
scheduler=scheduler,
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
max_noise_level=350,
)
sd_pipe = sd_pipe.to(device)
sd_pipe.set_progress_bar_config(disable=None)
prompt = "A painting of a squirrel eating a burger"
output = sd_pipe(
2 * [prompt],
image=2 * [low_res_image],
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
)
image = output.images
assert image.shape[0] == 2
generator = torch.Generator(device=device).manual_seed(0)
output = sd_pipe(
[prompt],
image=low_res_image,
generator=generator,
num_images_per_prompt=2,
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
)
image = output.images
assert image.shape[0] == 2
def test_stable_diffusion_upscale_prompt_embeds(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
unet = self.dummy_cond_unet_upscale
low_res_scheduler = DDPMScheduler()
scheduler = DDIMScheduler(prediction_type="v_prediction")
vae = self.dummy_vae
text_encoder = self.dummy_text_encoder
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
# make sure here that pndm scheduler skips prk
sd_pipe = StableDiffusionUpscalePipeline(
unet=unet,
low_res_scheduler=low_res_scheduler,
scheduler=scheduler,
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
max_noise_level=350,
)
sd_pipe = sd_pipe.to(device)
sd_pipe.set_progress_bar_config(disable=None)
prompt = "A painting of a squirrel eating a burger"
generator = torch.Generator(device=device).manual_seed(0)
output = sd_pipe(
[prompt],
image=low_res_image,
generator=generator,
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
)
image = output.images
generator = torch.Generator(device=device).manual_seed(0)
prompt_embeds, negative_prompt_embeds = sd_pipe.encode_prompt(prompt, device, 1, False)
if negative_prompt_embeds is not None:
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
image_from_prompt_embeds = sd_pipe(
prompt_embeds=prompt_embeds,
image=[low_res_image],
generator=generator,
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
return_dict=False,
)[0]
image_slice = image[0, -3:, -3:, -1]
image_from_prompt_embeds_slice = image_from_prompt_embeds[0, -3:, -3:, -1]
expected_height_width = low_res_image.size[0] * 4
assert image.shape == (1, expected_height_width, expected_height_width, 3)
expected_slice = np.array([0.3113, 0.3910, 0.4272, 0.4859, 0.5061, 0.4652, 0.5362, 0.5715, 0.5661])
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
assert np.abs(image_from_prompt_embeds_slice.flatten() - expected_slice).max() < 1e-2
@unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
def test_stable_diffusion_upscale_fp16(self):
"""Test that stable diffusion upscale works with fp16"""
unet = self.dummy_cond_unet_upscale
low_res_scheduler = DDPMScheduler()
scheduler = DDIMScheduler(prediction_type="v_prediction")
vae = self.dummy_vae
text_encoder = self.dummy_text_encoder
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
# put models in fp16, except vae as it overflows in fp16
unet = unet.half()
text_encoder = text_encoder.half()
# make sure here that pndm scheduler skips prk
sd_pipe = StableDiffusionUpscalePipeline(
unet=unet,
low_res_scheduler=low_res_scheduler,
scheduler=scheduler,
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
max_noise_level=350,
)
sd_pipe = sd_pipe.to(torch_device)
sd_pipe.set_progress_bar_config(disable=None)
prompt = "A painting of a squirrel eating a burger"
generator = torch.manual_seed(0)
image = sd_pipe(
[prompt],
image=low_res_image,
generator=generator,
num_inference_steps=2,
output_type="np",
).images
expected_height_width = low_res_image.size[0] * 4
assert image.shape == (1, expected_height_width, expected_height_width, 3)
def test_stable_diffusion_upscale_from_save_pretrained(self):
pipes = []
device = "cpu" # ensure determinism for the device-dependent torch.Generator
low_res_scheduler = DDPMScheduler()
scheduler = DDIMScheduler(prediction_type="v_prediction")
tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
# make sure here that pndm scheduler skips prk
sd_pipe = StableDiffusionUpscalePipeline(
unet=self.dummy_cond_unet_upscale,
low_res_scheduler=low_res_scheduler,
scheduler=scheduler,
vae=self.dummy_vae,
text_encoder=self.dummy_text_encoder,
tokenizer=tokenizer,
max_noise_level=350,
)
sd_pipe = sd_pipe.to(device)
pipes.append(sd_pipe)
with tempfile.TemporaryDirectory() as tmpdirname:
sd_pipe.save_pretrained(tmpdirname)
sd_pipe = StableDiffusionUpscalePipeline.from_pretrained(tmpdirname).to(device)
pipes.append(sd_pipe)
prompt = "A painting of a squirrel eating a burger"
image = self.dummy_image.cpu().permute(0, 2, 3, 1)[0]
low_res_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
image_slices = []
for pipe in pipes:
generator = torch.Generator(device=device).manual_seed(0)
image = pipe(
[prompt],
image=low_res_image,
generator=generator,
guidance_scale=6.0,
noise_level=20,
num_inference_steps=2,
output_type="np",
).images
image_slices.append(image[0, -3:, -3:, -1].flatten())
assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
@slow
@require_torch_gpu
class StableDiffusionUpscalePipelineIntegrationTests(unittest.TestCase):
def setUp(self):
# clean up the VRAM before each test
super().setUp()
gc.collect()
torch.cuda.empty_cache()
def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
def test_stable_diffusion_upscale_pipeline(self):
image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
"/sd2-upscale/low_res_cat.png"
)
expected_image = load_numpy(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale"
"/upsampled_cat.npy"
)
model_id = "stabilityai/stable-diffusion-x4-upscaler"
pipe = StableDiffusionUpscalePipeline.from_pretrained(model_id)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing()
prompt = "a cat sitting on a park bench"
generator = torch.manual_seed(0)
output = pipe(
prompt=prompt,
image=image,
generator=generator,
output_type="np",
)
image = output.images[0]
assert image.shape == (512, 512, 3)
assert np.abs(expected_image - image).max() < 1e-3
def test_stable_diffusion_upscale_pipeline_fp16(self):
image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
"/sd2-upscale/low_res_cat.png"
)
expected_image = load_numpy(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale"
"/upsampled_cat_fp16.npy"
)
model_id = "stabilityai/stable-diffusion-x4-upscaler"
pipe = StableDiffusionUpscalePipeline.from_pretrained(
model_id,
torch_dtype=torch.float16,
)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing()
prompt = "a cat sitting on a park bench"
generator = torch.manual_seed(0)
output = pipe(
prompt=prompt,
image=image,
generator=generator,
output_type="np",
)
image = output.images[0]
assert image.shape == (512, 512, 3)
assert np.abs(expected_image - image).max() < 5e-1
def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
torch.cuda.reset_peak_memory_stats()
image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
"/sd2-upscale/low_res_cat.png"
)
model_id = "stabilityai/stable-diffusion-x4-upscaler"
pipe = StableDiffusionUpscalePipeline.from_pretrained(
model_id,
torch_dtype=torch.float16,
)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing(1)
pipe.enable_sequential_cpu_offload()
prompt = "a cat sitting on a park bench"
generator = torch.manual_seed(0)
_ = pipe(
prompt=prompt,
image=image,
generator=generator,
num_inference_steps=5,
output_type="np",
)
mem_bytes = torch.cuda.max_memory_allocated()
# make sure that less than 2.9 GB is allocated
assert mem_bytes < 2.9 * 10**9
|