Spaces:
Running
on
A10G
Running
on
A10G
File size: 9,993 Bytes
f1069cc |
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 |
# coding=utf-8
# Copyright 2023 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 unittest
import numpy as np
import torch
from transformers import XLMRobertaTokenizer
from diffusers import (
AltDiffusionImg2ImgPipeline,
AutoencoderKL,
PNDMScheduler,
UNet2DConditionModel,
)
from diffusers.image_processor import VaeImageProcessor
from diffusers.pipelines.alt_diffusion.modeling_roberta_series import (
RobertaSeriesConfig,
RobertaSeriesModelWithTransformation,
)
from diffusers.utils import floats_tensor, load_image, load_numpy, slow, torch_device
from diffusers.utils.testing_utils import require_torch_gpu
torch.backends.cuda.matmul.allow_tf32 = False
class AltDiffusionImg2ImgPipelineFastTests(unittest.TestCase):
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(self):
torch.manual_seed(0)
model = UNet2DConditionModel(
block_out_channels=(32, 64),
layers_per_block=2,
sample_size=32,
in_channels=4,
out_channels=4,
down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
cross_attention_dim=32,
)
return model
@property
def dummy_vae(self):
torch.manual_seed(0)
model = AutoencoderKL(
block_out_channels=[32, 64],
in_channels=3,
out_channels=3,
down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
latent_channels=4,
)
return model
@property
def dummy_text_encoder(self):
torch.manual_seed(0)
config = RobertaSeriesConfig(
hidden_size=32,
project_dim=32,
intermediate_size=37,
layer_norm_eps=1e-05,
num_attention_heads=4,
num_hidden_layers=5,
pad_token_id=1,
vocab_size=5006,
)
return RobertaSeriesModelWithTransformation(config)
@property
def dummy_extractor(self):
def extract(*args, **kwargs):
class Out:
def __init__(self):
self.pixel_values = torch.ones([0])
def to(self, device):
self.pixel_values.to(device)
return self
return Out()
return extract
def test_stable_diffusion_img2img_default_case(self):
device = "cpu" # ensure determinism for the device-dependent torch.Generator
unet = self.dummy_cond_unet
scheduler = PNDMScheduler(skip_prk_steps=True)
vae = self.dummy_vae
bert = self.dummy_text_encoder
tokenizer = XLMRobertaTokenizer.from_pretrained("hf-internal-testing/tiny-xlm-roberta")
tokenizer.model_max_length = 77
init_image = self.dummy_image.to(device)
# make sure here that pndm scheduler skips prk
alt_pipe = AltDiffusionImg2ImgPipeline(
unet=unet,
scheduler=scheduler,
vae=vae,
text_encoder=bert,
tokenizer=tokenizer,
safety_checker=None,
feature_extractor=self.dummy_extractor,
)
alt_pipe.image_processor = VaeImageProcessor(vae_scale_factor=alt_pipe.vae_scale_factor, do_normalize=False)
alt_pipe = alt_pipe.to(device)
alt_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 = alt_pipe(
[prompt],
generator=generator,
guidance_scale=6.0,
num_inference_steps=2,
output_type="np",
image=init_image,
)
image = output.images
generator = torch.Generator(device=device).manual_seed(0)
image_from_tuple = alt_pipe(
[prompt],
generator=generator,
guidance_scale=6.0,
num_inference_steps=2,
output_type="np",
image=init_image,
return_dict=False,
)[0]
image_slice = image[0, -3:, -3:, -1]
image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]
assert image.shape == (1, 32, 32, 3)
expected_slice = np.array([0.4115, 0.3870, 0.4089, 0.4807, 0.4668, 0.4144, 0.4151, 0.4721, 0.4569])
assert np.abs(image_slice.flatten() - expected_slice).max() < 5e-3
assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 5e-3
@unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
def test_stable_diffusion_img2img_fp16(self):
"""Test that stable diffusion img2img works with fp16"""
unet = self.dummy_cond_unet
scheduler = PNDMScheduler(skip_prk_steps=True)
vae = self.dummy_vae
bert = self.dummy_text_encoder
tokenizer = XLMRobertaTokenizer.from_pretrained("hf-internal-testing/tiny-xlm-roberta")
tokenizer.model_max_length = 77
init_image = self.dummy_image.to(torch_device)
# put models in fp16
unet = unet.half()
vae = vae.half()
bert = bert.half()
# make sure here that pndm scheduler skips prk
alt_pipe = AltDiffusionImg2ImgPipeline(
unet=unet,
scheduler=scheduler,
vae=vae,
text_encoder=bert,
tokenizer=tokenizer,
safety_checker=None,
feature_extractor=self.dummy_extractor,
)
alt_pipe.image_processor = VaeImageProcessor(vae_scale_factor=alt_pipe.vae_scale_factor, do_normalize=False)
alt_pipe = alt_pipe.to(torch_device)
alt_pipe.set_progress_bar_config(disable=None)
prompt = "A painting of a squirrel eating a burger"
generator = torch.manual_seed(0)
image = alt_pipe(
[prompt],
generator=generator,
num_inference_steps=2,
output_type="np",
image=init_image,
).images
assert image.shape == (1, 32, 32, 3)
@unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
def test_stable_diffusion_img2img_pipeline_multiple_of_8(self):
init_image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
"/img2img/sketch-mountains-input.jpg"
)
# resize to resolution that is divisible by 8 but not 16 or 32
init_image = init_image.resize((760, 504))
model_id = "BAAI/AltDiffusion"
pipe = AltDiffusionImg2ImgPipeline.from_pretrained(
model_id,
safety_checker=None,
)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing()
prompt = "A fantasy landscape, trending on artstation"
generator = torch.manual_seed(0)
output = pipe(
prompt=prompt,
image=init_image,
strength=0.75,
guidance_scale=7.5,
generator=generator,
output_type="np",
)
image = output.images[0]
image_slice = image[255:258, 383:386, -1]
assert image.shape == (504, 760, 3)
expected_slice = np.array([0.9358, 0.9397, 0.9599, 0.9901, 1.0000, 1.0000, 0.9882, 1.0000, 1.0000])
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
@slow
@require_torch_gpu
class AltDiffusionImg2ImgPipelineIntegrationTests(unittest.TestCase):
def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
def test_stable_diffusion_img2img_pipeline_default(self):
init_image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
"/img2img/sketch-mountains-input.jpg"
)
init_image = init_image.resize((768, 512))
expected_image = load_numpy(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/fantasy_landscape_alt.npy"
)
model_id = "BAAI/AltDiffusion"
pipe = AltDiffusionImg2ImgPipeline.from_pretrained(
model_id,
safety_checker=None,
)
pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
pipe.enable_attention_slicing()
prompt = "A fantasy landscape, trending on artstation"
generator = torch.manual_seed(0)
output = pipe(
prompt=prompt,
image=init_image,
strength=0.75,
guidance_scale=7.5,
generator=generator,
output_type="np",
)
image = output.images[0]
assert image.shape == (512, 768, 3)
# img2img is flaky across GPUs even in fp32, so using MAE here
assert np.abs(expected_image - image).max() < 1e-3
|