test_latent_diffusion.py 6.97 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
#
# 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.

16
import gc
17
18
19
20
import unittest

import numpy as np
import torch
21
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
22
23

from diffusers import AutoencoderKL, DDIMScheduler, LDMTextToImagePipeline, UNet2DConditionModel
24
from diffusers.utils.testing_utils import load_numpy, nightly, require_torch_gpu, slow, torch_device
25

26
from ...pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
27
28
from ...test_pipelines_common import PipelineTesterMixin

29
30
31
32

torch.backends.cuda.matmul.allow_tf32 = False


33
34
class LDMTextToImagePipelineFastTests(PipelineTesterMixin, unittest.TestCase):
    pipeline_class = LDMTextToImagePipeline
35
36
37
38
39
40
41
42
43
44
45
46
    params = TEXT_TO_IMAGE_PARAMS - {
        "negative_prompt",
        "negative_prompt_embeds",
        "cross_attention_kwargs",
        "prompt_embeds",
    }
    required_optional_params = PipelineTesterMixin.required_optional_params - {
        "num_images_per_prompt",
        "callback",
        "callback_steps",
    }
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
47
48
49
    test_cpu_offload = False

    def get_dummy_components(self):
50
        torch.manual_seed(0)
51
        unet = UNet2DConditionModel(
52
53
54
55
56
57
58
59
60
            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,
        )
61
62
63
64
65
66
67
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
        )
68
        torch.manual_seed(0)
69
70
        vae = AutoencoderKL(
            block_out_channels=(32, 64),
71
72
            in_channels=3,
            out_channels=3,
73
74
            down_block_types=("DownEncoderBlock2D", "DownEncoderBlock2D"),
            up_block_types=("UpDecoderBlock2D", "UpDecoderBlock2D"),
75
76
77
            latent_channels=4,
        )
        torch.manual_seed(0)
78
        text_encoder_config = CLIPTextConfig(
79
80
81
82
83
84
85
86
87
88
            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,
        )
89
        text_encoder = CLIPTextModel(text_encoder_config)
90
91
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vqvae": vae,
            "bert": text_encoder,
            "tokenizer": tokenizer,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs
114
115

    def test_inference_text2img(self):
116
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
117

118
119
120
121
        components = self.get_dummy_components()
        pipe = LDMTextToImagePipeline(**components)
        pipe.to(device)
        pipe.set_progress_bar_config(disable=None)
122

123
124
        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
125
126
        image_slice = image[0, -3:, -3:, -1]

127
        assert image.shape == (1, 16, 16, 3)
128
        expected_slice = np.array([0.6101, 0.6156, 0.5622, 0.4895, 0.6661, 0.3804, 0.5748, 0.6136, 0.5014])
129

130
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
131

132

133
134
135
136
137
138
139
140
141
@slow
@require_torch_gpu
class LDMTextToImagePipelineSlowTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, device, dtype=torch.float32, seed=0):
142
        generator = torch.manual_seed(seed)
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
        latents = np.random.RandomState(seed).standard_normal((1, 4, 32, 32))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs

    def test_ldm_default_ddim(self):
        pipe = LDMTextToImagePipeline.from_pretrained("CompVis/ldm-text2im-large-256").to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
162
163

        assert image.shape == (1, 256, 256, 3)
164
165
166
167
168
169
170
171
172
173
174
175
176
177
        expected_slice = np.array([0.51825, 0.52850, 0.52543, 0.54258, 0.52304, 0.52569, 0.54363, 0.55276, 0.56878])
        max_diff = np.abs(expected_slice - image_slice).max()
        assert max_diff < 1e-3


@nightly
@require_torch_gpu
class LDMTextToImagePipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, device, dtype=torch.float32, seed=0):
178
        generator = torch.manual_seed(seed)
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
        latents = np.random.RandomState(seed).standard_normal((1, 4, 32, 32))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 50,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs

    def test_ldm_default_ddim(self):
        pipe = LDMTextToImagePipeline.from_pretrained("CompVis/ldm-text2im-large-256").to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/ldm_text2img/ldm_large_256_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3