File size: 12,951 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
# coding=utf-8
# Copyright 2024 Harutatsu Akiyama, Jinbin Bai, and 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 random
import unittest

import numpy as np
import torch
from PIL import Image
from transformers import (
    CLIPImageProcessor,
    CLIPTextConfig,
    CLIPTextModel,
    CLIPTextModelWithProjection,
    CLIPTokenizer,
    CLIPVisionConfig,
    CLIPVisionModelWithProjection,
)

from diffusers import (
    AutoencoderKL,
    ControlNetModel,
    EulerDiscreteScheduler,
    StableDiffusionXLControlNetInpaintPipeline,
    UNet2DConditionModel,
)
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.testing_utils import enable_full_determinism, floats_tensor, require_torch_gpu, torch_device

from ..pipeline_params import (
    IMAGE_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_BATCH_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
    TEXT_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_PARAMS,
)
from ..test_pipelines_common import (
    PipelineKarrasSchedulerTesterMixin,
    PipelineLatentTesterMixin,
    PipelineTesterMixin,
)


enable_full_determinism()


class ControlNetPipelineSDXLFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
    pipeline_class = StableDiffusionXLControlNetInpaintPipeline
    params = TEXT_TO_IMAGE_PARAMS
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
    image_params = frozenset(IMAGE_TO_IMAGE_IMAGE_PARAMS.union({"mask_image", "control_image"}))
    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union(
        {
            "add_text_embeds",
            "add_time_ids",
            "mask",
            "masked_image_latents",
        }
    )

    def get_dummy_components(self):
        torch.manual_seed(0)
        unet = 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"),
            # SD2-specific config below
            attention_head_dim=(2, 4),
            use_linear_projection=True,
            addition_embed_type="text_time",
            addition_time_embed_dim=8,
            transformer_layers_per_block=(1, 2),
            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
            cross_attention_dim=64,
        )
        torch.manual_seed(0)
        controlnet = ControlNetModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            in_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            conditioning_embedding_out_channels=(16, 32),
            # SD2-specific config below
            attention_head_dim=(2, 4),
            use_linear_projection=True,
            addition_embed_type="text_time",
            addition_time_embed_dim=8,
            transformer_layers_per_block=(1, 2),
            projection_class_embeddings_input_dim=80,  # 6 * 8 + 32
            cross_attention_dim=64,
        )
        scheduler = EulerDiscreteScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            steps_offset=1,
            beta_schedule="scaled_linear",
            timestep_spacing="leading",
        )
        torch.manual_seed(0)
        vae = 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,
        )
        torch.manual_seed(0)
        text_encoder_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=32,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        torch.manual_seed(0)
        text_encoder_2 = CLIPTextModelWithProjection(text_encoder_config)
        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        image_encoder_config = CLIPVisionConfig(
            hidden_size=32,
            image_size=224,
            projection_dim=32,
            intermediate_size=37,
            num_attention_heads=4,
            num_channels=3,
            num_hidden_layers=5,
            patch_size=14,
        )

        image_encoder = CLIPVisionModelWithProjection(image_encoder_config)

        feature_extractor = CLIPImageProcessor(
            crop_size=224,
            do_center_crop=True,
            do_normalize=True,
            do_resize=True,
            image_mean=[0.48145466, 0.4578275, 0.40821073],
            image_std=[0.26862954, 0.26130258, 0.27577711],
            resample=3,
            size=224,
        )

        components = {
            "unet": unet,
            "controlnet": controlnet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "text_encoder_2": text_encoder_2,
            "tokenizer_2": tokenizer_2,
            "image_encoder": image_encoder,
            "feature_extractor": feature_extractor,
        }
        return components

    def get_dummy_inputs(self, device, seed=0, img_res=64):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)

        # Get random floats in [0, 1] as image
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        image = image.cpu().permute(0, 2, 3, 1)[0]
        mask_image = torch.ones_like(image)
        controlnet_embedder_scale_factor = 2
        control_image = (
            floats_tensor(
                (1, 3, 32 * controlnet_embedder_scale_factor, 32 * controlnet_embedder_scale_factor),
                rng=random.Random(seed),
            )
            .to(device)
            .cpu()
        )
        control_image = control_image.cpu().permute(0, 2, 3, 1)[0]
        # Convert image and mask_image to [0, 255]
        image = 255 * image
        mask_image = 255 * mask_image
        control_image = 255 * control_image
        # Convert to PIL image
        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((img_res, img_res))
        mask_image = Image.fromarray(np.uint8(mask_image)).convert("L").resize((img_res, img_res))
        control_image = Image.fromarray(np.uint8(control_image)).convert("RGB").resize((img_res, img_res))

        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "np",
            "image": init_image,
            "mask_image": mask_image,
            "control_image": control_image,
        }
        return inputs

    def test_attention_slicing_forward_pass(self):
        return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)

    def test_dict_tuple_outputs_equivalent(self):
        expected_slice = None
        if torch_device == "cpu":
            expected_slice = np.array([0.5490, 0.5053, 0.4676, 0.5816, 0.5364, 0.4830, 0.5937, 0.5719, 0.4318])
        super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)

    @unittest.skipIf(
        torch_device != "cuda" or not is_xformers_available(),
        reason="XFormers attention is only available with CUDA and `xformers` installed",
    )
    def test_xformers_attention_forwardGenerator_pass(self):
        self._test_xformers_attention_forwardGenerator_pass(expected_max_diff=2e-3)

    def test_inference_batch_single_identical(self):
        self._test_inference_batch_single_identical(expected_max_diff=2e-3)

    @require_torch_gpu
    def test_stable_diffusion_xl_offloads(self):
        pipes = []
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(torch_device)
        pipes.append(sd_pipe)

        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe.enable_model_cpu_offload()
        pipes.append(sd_pipe)

        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe.enable_sequential_cpu_offload()
        pipes.append(sd_pipe)

        image_slices = []
        for pipe in pipes:
            pipe.unet.set_default_attn_processor()

            inputs = self.get_dummy_inputs(torch_device)
            image = pipe(**inputs).images

            image_slices.append(image[0, -3:, -3:, -1].flatten())

        assert np.abs(image_slices[0] - image_slices[1]).max() < 1e-3
        assert np.abs(image_slices[0] - image_slices[2]).max() < 1e-3

    def test_stable_diffusion_xl_multi_prompts(self):
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(torch_device)

        # forward with single prompt
        inputs = self.get_dummy_inputs(torch_device)
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        # forward with same prompt duplicated
        inputs = self.get_dummy_inputs(torch_device)
        inputs["prompt_2"] = inputs["prompt"]
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        # ensure the results are equal
        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

        # forward with different prompt
        inputs = self.get_dummy_inputs(torch_device)
        inputs["prompt_2"] = "different prompt"
        output = sd_pipe(**inputs)
        image_slice_3 = output.images[0, -3:, -3:, -1]

        # ensure the results are not equal
        assert np.abs(image_slice_1.flatten() - image_slice_3.flatten()).max() > 1e-4

        # manually set a negative_prompt
        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = "negative prompt"
        output = sd_pipe(**inputs)
        image_slice_1 = output.images[0, -3:, -3:, -1]

        # forward with same negative_prompt duplicated
        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = "negative prompt"
        inputs["negative_prompt_2"] = inputs["negative_prompt"]
        output = sd_pipe(**inputs)
        image_slice_2 = output.images[0, -3:, -3:, -1]

        # ensure the results are equal
        assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4

        # forward with different negative_prompt
        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt"] = "negative prompt"
        inputs["negative_prompt_2"] = "different negative prompt"
        output = sd_pipe(**inputs)
        image_slice_3 = output.images[0, -3:, -3:, -1]

        # ensure the results are not equal
        assert np.abs(image_slice_1.flatten() - image_slice_3.flatten()).max() > 1e-4

    def test_controlnet_sdxl_guess(self):
        device = "cpu"

        components = self.get_dummy_components()

        sd_pipe = self.pipeline_class(**components)
        sd_pipe = sd_pipe.to(device)

        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        inputs["guess_mode"] = True

        output = sd_pipe(**inputs)
        image_slice = output.images[0, -3:, -3:, -1]
        expected_slice = np.array([0.549, 0.5053, 0.4676, 0.5816, 0.5364, 0.483, 0.5937, 0.5719, 0.4318])

        # make sure that it's equal
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-4

    # TODO(Patrick, Sayak) - skip for now as this requires more refiner tests
    def test_save_load_optional_components(self):
        pass

    def test_float16_inference(self):
        super().test_float16_inference(expected_max_diff=5e-1)