test_cycle_diffusion.py 9.63 KB
Newer Older
1
# coding=utf-8
Patrick von Platen's avatar
Patrick von Platen committed
2
# Copyright 2023 HuggingFace Inc.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
#
# 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
22
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
23

24
from diffusers import AutoencoderKL, CycleDiffusionPipeline, DDIMScheduler, UNet2DConditionModel
Dhruv Nair's avatar
Dhruv Nair committed
25
26
27
28
29
from diffusers.utils.testing_utils import (
    enable_full_determinism,
    floats_tensor,
    load_image,
    load_numpy,
30
    nightly,
Dhruv Nair's avatar
Dhruv Nair committed
31
32
33
34
    require_torch_gpu,
    skip_mps,
    torch_device,
)
35

36
37
38
39
40
from ..pipeline_params import (
    IMAGE_TO_IMAGE_IMAGE_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_PARAMS,
)
41
from ..test_pipelines_common import PipelineLatentTesterMixin, PipelineTesterMixin
42
43


44
enable_full_determinism()
45
46


47
class CycleDiffusionPipelineFastTests(PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase):
48
    pipeline_class = CycleDiffusionPipeline
49
50
51
52
53
54
55
    params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS - {
        "negative_prompt",
        "height",
        "width",
        "negative_prompt_embeds",
    }
    required_optional_params = PipelineTesterMixin.required_optional_params - {"latents"}
56
    batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS.union({"source_prompt"})
57
58
    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
    image_latents_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
59

60
    def get_dummy_components(self):
61
        torch.manual_seed(0)
62
        unet = UNet2DConditionModel(
63
64
65
66
67
68
69
70
71
            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,
        )
72
73
74
75
76
77
78
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            num_train_timesteps=1000,
            clip_sample=False,
            set_alpha_to_one=False,
79
80
        )
        torch.manual_seed(0)
81
        vae = AutoencoderKL(
82
83
84
85
86
87
88
89
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
        )
        torch.manual_seed(0)
90
        text_encoder_config = CLIPTextConfig(
91
92
93
94
95
96
97
98
99
100
            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,
        )
101
102
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
103

104
105
106
107
108
109
110
111
112
113
114
115
116
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
117
        image = image / 2 + 0.5
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "An astronaut riding an elephant",
            "source_prompt": "An astronaut riding a horse",
            "image": image,
            "generator": generator,
            "num_inference_steps": 2,
            "eta": 0.1,
            "strength": 0.8,
            "guidance_scale": 3,
            "source_guidance_scale": 1,
            "output_type": "numpy",
        }
        return inputs
135
136
137
138

    def test_stable_diffusion_cycle(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

139
140
141
142
        components = self.get_dummy_components()
        pipe = CycleDiffusionPipeline(**components)
        pipe = pipe.to(device)
        pipe.set_progress_bar_config(disable=None)
143

144
145
        inputs = self.get_dummy_inputs(device)
        output = pipe(**inputs)
146
147
148
149
150
151
152
153
154
155
156
        images = output.images

        image_slice = images[0, -3:, -3:, -1]

        assert images.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.4459, 0.4943, 0.4544, 0.6643, 0.5474, 0.4327, 0.5701, 0.5959, 0.5179])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
    def test_stable_diffusion_cycle_fp16(self):
157
158
159
160
161
162
163
164
165
166
        components = self.get_dummy_components()
        for name, module in components.items():
            if hasattr(module, "half"):
                components[name] = module.half()
        pipe = CycleDiffusionPipeline(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output = pipe(**inputs)
167
168
169
170
171
172
173
174
175
        images = output.images

        image_slice = images[0, -3:, -3:, -1]

        assert images.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.3506, 0.4543, 0.446, 0.4575, 0.5195, 0.4155, 0.5273, 0.518, 0.4116])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
    @skip_mps
    def test_save_load_local(self):
        return super().test_save_load_local()

    @unittest.skip("non-deterministic pipeline")
    def test_inference_batch_single_identical(self):
        return super().test_inference_batch_single_identical()

    @skip_mps
    def test_dict_tuple_outputs_equivalent(self):
        return super().test_dict_tuple_outputs_equivalent()

    @skip_mps
    def test_save_load_optional_components(self):
        return super().test_save_load_optional_components()

    @skip_mps
    def test_attention_slicing_forward_pass(self):
        return super().test_attention_slicing_forward_pass()

196

197
@nightly
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
@require_torch_gpu
class CycleDiffusionPipelineIntegrationTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def test_cycle_diffusion_pipeline_fp16(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/cycle-diffusion/black_colored_car.png"
        )
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/cycle-diffusion/blue_colored_car_fp16.npy"
        )
        init_image = init_image.resize((512, 512))

        model_id = "CompVis/stable-diffusion-v1-4"
217
        scheduler = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler")
218
219
220
221
222
223
224
225
226
227
228
        pipe = CycleDiffusionPipeline.from_pretrained(
            model_id, scheduler=scheduler, safety_checker=None, torch_dtype=torch.float16, revision="fp16"
        )

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        source_prompt = "A black colored car"
        prompt = "A blue colored car"

229
        generator = torch.manual_seed(0)
230
231
232
        output = pipe(
            prompt=prompt,
            source_prompt=source_prompt,
233
            image=init_image,
234
235
236
237
238
            num_inference_steps=100,
            eta=0.1,
            strength=0.85,
            guidance_scale=3,
            source_guidance_scale=1,
Patrick von Platen's avatar
Patrick von Platen committed
239
            generator=generator,
240
241
242
243
244
            output_type="np",
        )
        image = output.images

        # the values aren't exactly equal, but the images look the same visually
Patrick von Platen's avatar
Patrick von Platen committed
245
        assert np.abs(image - expected_image).max() < 5e-1
246
247
248
249
250
251
252
253
254
255
256
257

    def test_cycle_diffusion_pipeline(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/cycle-diffusion/black_colored_car.png"
        )
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/cycle-diffusion/blue_colored_car.npy"
        )
        init_image = init_image.resize((512, 512))

        model_id = "CompVis/stable-diffusion-v1-4"
258
        scheduler = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler")
259
260
261
262
263
264
265
266
267
        pipe = CycleDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, safety_checker=None)

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        source_prompt = "A black colored car"
        prompt = "A blue colored car"

268
        generator = torch.manual_seed(0)
269
270
271
        output = pipe(
            prompt=prompt,
            source_prompt=source_prompt,
272
            image=init_image,
273
274
275
276
277
            num_inference_steps=100,
            eta=0.1,
            strength=0.85,
            guidance_scale=3,
            source_guidance_scale=1,
Patrick von Platen's avatar
Patrick von Platen committed
278
            generator=generator,
279
280
281
282
            output_type="np",
        )
        image = output.images

283
        assert np.abs(image - expected_image).max() < 2e-2