File size: 16,682 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
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
# coding=utf-8
# Copyright 2024 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 PIL import Image
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer

from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    EulerAncestralDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionInstructPix2PixPipeline,
    UNet2DConditionModel,
)
from diffusers.image_processor import VaeImageProcessor
from diffusers.utils.testing_utils import (
    enable_full_determinism,
    floats_tensor,
    load_image,
    require_torch_gpu,
    slow,
    torch_device,
)

from ..pipeline_params import (
    IMAGE_TO_IMAGE_IMAGE_PARAMS,
    TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
)
from ..test_pipelines_common import (
    PipelineKarrasSchedulerTesterMixin,
    PipelineLatentTesterMixin,
    PipelineTesterMixin,
)


enable_full_determinism()


class StableDiffusionInstructPix2PixPipelineFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
    pipeline_class = StableDiffusionInstructPix2PixPipeline
    params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS - {"height", "width", "cross_attention_kwargs"}
    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
    image_latents_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union({"image_latents"}) - {"negative_prompt_embeds"}

    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=8,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        scheduler = PNDMScheduler(skip_prk_steps=True)
        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,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
            "image_encoder": None,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        image = image.cpu().permute(0, 2, 3, 1)[0]
        image = Image.fromarray(np.uint8(image)).convert("RGB")
        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",
            "image": image,
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "image_guidance_scale": 1,
            "output_type": "np",
        }
        return inputs

    def test_stable_diffusion_pix2pix_default_case(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInstructPix2PixPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]
        assert image.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.7526, 0.3750, 0.4547, 0.6117, 0.5866, 0.5016, 0.4327, 0.5642, 0.4815])

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

    def test_stable_diffusion_pix2pix_negative_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInstructPix2PixPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        negative_prompt = "french fries"
        output = sd_pipe(**inputs, negative_prompt=negative_prompt)
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.7511, 0.3642, 0.4553, 0.6236, 0.5797, 0.5013, 0.4343, 0.5611, 0.4831])

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

    def test_stable_diffusion_pix2pix_multiple_init_images(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInstructPix2PixPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        inputs["prompt"] = [inputs["prompt"]] * 2

        image = np.array(inputs["image"]).astype(np.float32) / 255.0
        image = torch.from_numpy(image).unsqueeze(0).to(device)
        image = image / 2 + 0.5
        image = image.permute(0, 3, 1, 2)
        inputs["image"] = image.repeat(2, 1, 1, 1)

        image = sd_pipe(**inputs).images
        image_slice = image[-1, -3:, -3:, -1]

        assert image.shape == (2, 32, 32, 3)
        expected_slice = np.array([0.5812, 0.5748, 0.5222, 0.5908, 0.5695, 0.7174, 0.6804, 0.5523, 0.5579])

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

    def test_stable_diffusion_pix2pix_euler(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        components["scheduler"] = EulerAncestralDiscreteScheduler(
            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
        )
        sd_pipe = StableDiffusionInstructPix2PixPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        slice = [round(x, 4) for x in image_slice.flatten().tolist()]
        print(",".join([str(x) for x in slice]))

        assert image.shape == (1, 32, 32, 3)
        expected_slice = np.array([0.7417, 0.3842, 0.4732, 0.5776, 0.5891, 0.5139, 0.4052, 0.5673, 0.4986])

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

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

    # Overwrite the default test_latents_inputs because pix2pix encode the image differently
    def test_latents_input(self):
        components = self.get_dummy_components()
        pipe = StableDiffusionInstructPix2PixPipeline(**components)
        pipe.image_processor = VaeImageProcessor(do_resize=False, do_normalize=False)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        out = pipe(**self.get_dummy_inputs_by_type(torch_device, input_image_type="pt"))[0]

        vae = components["vae"]
        inputs = self.get_dummy_inputs_by_type(torch_device, input_image_type="pt")

        for image_param in self.image_latents_params:
            if image_param in inputs.keys():
                inputs[image_param] = vae.encode(inputs[image_param]).latent_dist.mode()

        out_latents_inputs = pipe(**inputs)[0]

        max_diff = np.abs(out - out_latents_inputs).max()
        self.assertLess(max_diff, 1e-4, "passing latents as image input generate different result from passing image")

    # Override the default test_callback_cfg because pix2pix create inputs for cfg differently
    def test_callback_cfg(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        def callback_no_cfg(pipe, i, t, callback_kwargs):
            if i == 1:
                for k, w in callback_kwargs.items():
                    if k in self.callback_cfg_params:
                        callback_kwargs[k] = callback_kwargs[k].chunk(3)[0]
                pipe._guidance_scale = 1.0

            return callback_kwargs

        inputs = self.get_dummy_inputs(torch_device)
        inputs["guidance_scale"] = 1.0
        inputs["num_inference_steps"] = 2
        out_no_cfg = pipe(**inputs)[0]

        inputs["guidance_scale"] = 7.5
        inputs["callback_on_step_end"] = callback_no_cfg
        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
        out_callback_no_cfg = pipe(**inputs)[0]

        assert out_no_cfg.shape == out_callback_no_cfg.shape


@slow
@require_torch_gpu
class StableDiffusionInstructPix2PixPipelineSlowTests(unittest.TestCase):
    def setUp(self):
        super().setUp()
        gc.collect()
        torch.cuda.empty_cache()

    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, seed=0):
        generator = torch.manual_seed(seed)
        image = load_image(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_pix2pix/example.jpg"
        )
        inputs = {
            "prompt": "turn him into a cyborg",
            "image": image,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "image_guidance_scale": 1.0,
            "output_type": "np",
        }
        return inputs

    def test_stable_diffusion_pix2pix_default(self):
        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            "timbrooks/instruct-pix2pix", safety_checker=None
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

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

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.5902, 0.6015, 0.6027, 0.5983, 0.6092, 0.6061, 0.5765, 0.5785, 0.5555])

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

    def test_stable_diffusion_pix2pix_k_lms(self):
        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            "timbrooks/instruct-pix2pix", safety_checker=None
        )
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

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

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.6578, 0.6817, 0.6972, 0.6761, 0.6856, 0.6916, 0.6428, 0.6516, 0.6301])

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

    def test_stable_diffusion_pix2pix_ddim(self):
        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            "timbrooks/instruct-pix2pix", safety_checker=None
        )
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

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

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.3828, 0.3834, 0.3818, 0.3792, 0.3865, 0.3752, 0.3792, 0.3847, 0.3753])

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

    def test_stable_diffusion_pix2pix_intermediate_state(self):
        number_of_steps = 0

        def callback_fn(step: int, timestep: int, latents: torch.Tensor) -> None:
            callback_fn.has_been_called = True
            nonlocal number_of_steps
            number_of_steps += 1
            if step == 1:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array([-0.2463, -0.4644, -0.9756, 1.5176, 1.4414, 0.7866, 0.9897, 0.8521, 0.7983])

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
            elif step == 2:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array([-0.2644, -0.4626, -0.9653, 1.5176, 1.4551, 0.7686, 0.9805, 0.8452, 0.8115])

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2

        callback_fn.has_been_called = False

        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            "timbrooks/instruct-pix2pix", safety_checker=None, torch_dtype=torch.float16
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs()
        pipe(**inputs, callback=callback_fn, callback_steps=1)
        assert callback_fn.has_been_called
        assert number_of_steps == 3

    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            "timbrooks/instruct-pix2pix", safety_checker=None, torch_dtype=torch.float16
        )
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

        inputs = self.get_inputs()
        _ = pipe(**inputs)

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.2 GB is allocated
        assert mem_bytes < 2.2 * 10**9

    def test_stable_diffusion_pix2pix_pipeline_multiple_of_8(self):
        inputs = self.get_inputs()
        # resize to resolution that is divisible by 8 but not 16 or 32
        inputs["image"] = inputs["image"].resize((504, 504))

        model_id = "timbrooks/instruct-pix2pix"
        pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            model_id,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        output = pipe(**inputs)
        image = output.images[0]

        image_slice = image[255:258, 383:386, -1]

        assert image.shape == (504, 504, 3)
        expected_slice = np.array([0.2726, 0.2529, 0.2664, 0.2655, 0.2641, 0.2642, 0.2591, 0.2649, 0.2590])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 5e-3