File size: 7,886 Bytes
fd43906
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# 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 unittest

import numpy as np
import torch
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer

from diffusers import Transformer2DModel, VQDiffusionPipeline, VQDiffusionScheduler, VQModel
from diffusers.pipelines.vq_diffusion.pipeline_vq_diffusion import LearnedClassifierFreeSamplingEmbeddings
from diffusers.utils import load_numpy, slow, torch_device
from diffusers.utils.testing_utils import require_torch_gpu


torch.backends.cuda.matmul.allow_tf32 = False


class VQDiffusionPipelineFastTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    @property
    def num_embed(self):
        return 12

    @property
    def num_embeds_ada_norm(self):
        return 12

    @property
    def text_embedder_hidden_size(self):
        return 32

    @property
    def dummy_vqvae(self):
        torch.manual_seed(0)
        model = VQModel(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=3,
            num_vq_embeddings=self.num_embed,
            vq_embed_dim=3,
        )
        return model

    @property
    def dummy_tokenizer(self):
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
        return tokenizer

    @property
    def dummy_text_encoder(self):
        torch.manual_seed(0)
        config = CLIPTextConfig(
            bos_token_id=0,
            eos_token_id=2,
            hidden_size=self.text_embedder_hidden_size,
            intermediate_size=37,
            layer_norm_eps=1e-05,
            num_attention_heads=4,
            num_hidden_layers=5,
            pad_token_id=1,
            vocab_size=1000,
        )
        return CLIPTextModel(config)

    @property
    def dummy_transformer(self):
        torch.manual_seed(0)

        height = 12
        width = 12

        model_kwargs = {
            "attention_bias": True,
            "cross_attention_dim": 32,
            "attention_head_dim": height * width,
            "num_attention_heads": 1,
            "num_vector_embeds": self.num_embed,
            "num_embeds_ada_norm": self.num_embeds_ada_norm,
            "norm_num_groups": 32,
            "sample_size": width,
            "activation_fn": "geglu-approximate",
        }

        model = Transformer2DModel(**model_kwargs)
        return model

    def test_vq_diffusion(self):
        device = "cpu"

        vqvae = self.dummy_vqvae
        text_encoder = self.dummy_text_encoder
        tokenizer = self.dummy_tokenizer
        transformer = self.dummy_transformer
        scheduler = VQDiffusionScheduler(self.num_embed)
        learned_classifier_free_sampling_embeddings = LearnedClassifierFreeSamplingEmbeddings(learnable=False)

        pipe = VQDiffusionPipeline(
            vqvae=vqvae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            transformer=transformer,
            scheduler=scheduler,
            learned_classifier_free_sampling_embeddings=learned_classifier_free_sampling_embeddings,
        )
        pipe = pipe.to(device)
        pipe.set_progress_bar_config(disable=None)

        prompt = "teddy bear playing in the pool"

        generator = torch.Generator(device=device).manual_seed(0)
        output = pipe([prompt], generator=generator, num_inference_steps=2, output_type="np")
        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = pipe(
            [prompt], generator=generator, output_type="np", return_dict=False, num_inference_steps=2
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 24, 24, 3)

        expected_slice = np.array([0.6583, 0.6410, 0.5325, 0.5635, 0.5563, 0.4234, 0.6008, 0.5491, 0.4880])

        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_vq_diffusion_classifier_free_sampling(self):
        device = "cpu"

        vqvae = self.dummy_vqvae
        text_encoder = self.dummy_text_encoder
        tokenizer = self.dummy_tokenizer
        transformer = self.dummy_transformer
        scheduler = VQDiffusionScheduler(self.num_embed)
        learned_classifier_free_sampling_embeddings = LearnedClassifierFreeSamplingEmbeddings(
            learnable=True, hidden_size=self.text_embedder_hidden_size, length=tokenizer.model_max_length
        )

        pipe = VQDiffusionPipeline(
            vqvae=vqvae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            transformer=transformer,
            scheduler=scheduler,
            learned_classifier_free_sampling_embeddings=learned_classifier_free_sampling_embeddings,
        )
        pipe = pipe.to(device)
        pipe.set_progress_bar_config(disable=None)

        prompt = "teddy bear playing in the pool"

        generator = torch.Generator(device=device).manual_seed(0)
        output = pipe([prompt], generator=generator, num_inference_steps=2, output_type="np")
        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = pipe(
            [prompt], generator=generator, output_type="np", return_dict=False, num_inference_steps=2
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 24, 24, 3)

        expected_slice = np.array([0.6647, 0.6531, 0.5303, 0.5891, 0.5726, 0.4439, 0.6304, 0.5564, 0.4912])

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


@slow
@require_torch_gpu
class VQDiffusionPipelineIntegrationTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def test_vq_diffusion_classifier_free_sampling(self):
        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/vq_diffusion/teddy_bear_pool_classifier_free_sampling.npy"
        )

        pipeline = VQDiffusionPipeline.from_pretrained("microsoft/vq-diffusion-ithq")
        pipeline = pipeline.to(torch_device)
        pipeline.set_progress_bar_config(disable=None)

        # requires GPU generator for gumbel softmax
        # don't use GPU generator in tests though
        generator = torch.Generator(device=torch_device).manual_seed(0)
        output = pipeline(
            "teddy bear playing in the pool",
            num_images_per_prompt=1,
            generator=generator,
            output_type="np",
        )

        image = output.images[0]

        assert image.shape == (256, 256, 3)
        assert np.abs(expected_image - image).max() < 1e-2