test_stable_diffusion_inpaint.py 51.9 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
#
# 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
18
import traceback
19
20
21
22
import unittest

import numpy as np
import torch
23
from huggingface_hub import hf_hub_download
24
25
from PIL import Image
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
26
27

from diffusers import (
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
28
    AsymmetricAutoencoderKL,
29
    AutoencoderKL,
30
    DDIMScheduler,
31
    DPMSolverMultistepScheduler,
32
    LMSDiscreteScheduler,
33
34
35
36
    PNDMScheduler,
    StableDiffusionInpaintPipeline,
    UNet2DConditionModel,
)
37
from diffusers.models.attention_processor import AttnProcessor
38
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image
39
from diffusers.utils import floats_tensor, load_image, load_numpy, nightly, slow, torch_device
40
41
42
43
44
45
from diffusers.utils.testing_utils import (
    enable_full_determinism,
    require_torch_2,
    require_torch_gpu,
    run_test_in_subprocess,
)
46

47
from ...models.test_models_unet_2d_condition import create_lora_layers
48
from ..pipeline_params import TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS, TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
49
from ..test_pipelines_common import PipelineKarrasSchedulerTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin
50

51

52
enable_full_determinism()
53
54


55
56
57
58
59
60
61
62
63
64
65
66
# Will be run via run_test_in_subprocess
def _test_inpaint_compile(in_queue, out_queue, timeout):
    error = None
    try:
        inputs = in_queue.get(timeout=timeout)
        torch_device = inputs.pop("torch_device")
        seed = inputs.pop("seed")
        inputs["generator"] = torch.Generator(device=torch_device).manual_seed(seed)

        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
67
        pipe.unet.set_default_attn_processor()
68
69
70
71
72
73
74
75
76
77
78
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        pipe.unet.to(memory_format=torch.channels_last)
        pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
79
        expected_slice = np.array([0.0689, 0.0699, 0.0790, 0.0536, 0.0470, 0.0488, 0.041, 0.0508, 0.04179])
80
81
82
83
84
85
86
87
88
        assert np.abs(expected_slice - image_slice).max() < 3e-3
    except Exception:
        error = f"{traceback.format_exc()}"

    results = {"error": error}
    out_queue.put(results, timeout=timeout)
    out_queue.join()


89
90
91
class StableDiffusionInpaintPipelineFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
92
    pipeline_class = StableDiffusionInpaintPipeline
93
94
    params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
95
96
    image_params = frozenset([])
    # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess
97
    image_latents_params = frozenset([])
98

99
    def get_dummy_components(self):
100
        torch.manual_seed(0)
101
        unet = UNet2DConditionModel(
102
103
104
105
106
107
108
109
110
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=9,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
111
        scheduler = PNDMScheduler(skip_prk_steps=True)
112
        torch.manual_seed(0)
113
        vae = AutoencoderKL(
114
115
116
117
118
119
120
121
            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)
122
        text_encoder_config = CLIPTextConfig(
123
124
125
126
127
128
129
130
131
132
            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,
        )
133
        text_encoder = CLIPTextModel(text_encoder_config)
134
135
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

136
137
138
139
140
141
142
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
143
            "feature_extractor": None,
144
145
146
147
148
149
150
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        # TODO: use tensor inputs instead of PIL, this is here just to leave the old expected_slices untouched
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        image = image.cpu().permute(0, 2, 3, 1)[0]
Patrick von Platen's avatar
Patrick von Platen committed
151
152
        init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((64, 64))
        mask_image = Image.fromarray(np.uint8(image + 4)).convert("RGB").resize((64, 64))
153
154
155
156
157
158
159
160
161
162
163
164
165
166
        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": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs
167

168
169
170
171
    def test_stable_diffusion_inpaint(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
172
173
174
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

175
176
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
177
178
        image_slice = image[0, -3:, -3:, -1]

179
        assert image.shape == (1, 64, 64, 3)
Patrick von Platen's avatar
Patrick von Platen committed
180
        expected_slice = np.array([0.4723, 0.5731, 0.3939, 0.5441, 0.5922, 0.4392, 0.5059, 0.4651, 0.4474])
181

182
183
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

184
185
    def test_stable_diffusion_inpaint_image_tensor(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
186
187
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
188
189
190
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

191
192
193
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
        out_pil = output.images
194

195
196
197
198
199
        inputs = self.get_dummy_inputs(device)
        inputs["image"] = torch.tensor(np.array(inputs["image"]) / 127.5 - 1).permute(2, 0, 1).unsqueeze(0)
        inputs["mask_image"] = torch.tensor(np.array(inputs["mask_image"]) / 255).permute(2, 0, 1)[:1].unsqueeze(0)
        output = sd_pipe(**inputs)
        out_tensor = output.images
200

201
202
        assert out_pil.shape == (1, 64, 64, 3)
        assert np.abs(out_pil.flatten() - out_tensor.flatten()).max() < 5e-2
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
    def test_stable_diffusion_inpaint_lora(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator

        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

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

        # set lora layers
        lora_attn_procs = create_lora_layers(sd_pipe.unet)
        sd_pipe.unet.set_attn_processor(lora_attn_procs)
        sd_pipe = sd_pipe.to(torch_device)

        # forward 2
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs, cross_attention_kwargs={"scale": 0.0})
        image = output.images
        image_slice_1 = image[0, -3:, -3:, -1]

        # forward 3
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs, cross_attention_kwargs={"scale": 0.5})
        image = output.images
        image_slice_2 = image[0, -3:, -3:, -1]

        assert np.abs(image_slice - image_slice_1).max() < 1e-2
        assert np.abs(image_slice - image_slice_2).max() > 1e-2

238
239
240
    def test_inference_batch_single_identical(self):
        super().test_inference_batch_single_identical(expected_max_diff=3e-3)

241
242
243
244
245
246
247
248
249
250
251
252
253
254
    def test_stable_diffusion_inpaint_strength_zero_test(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)

        # check that the pipeline raises value error when num_inference_steps is < 1
        inputs["strength"] = 0.01
        with self.assertRaises(ValueError):
            sd_pipe(**inputs).images

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
    def test_stable_diffusion_inpaint_mask_latents(self):
        device = "cpu"
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        # normal mask + normal image
        ##  `image`: pil, `mask_image``: pil, `masked_image_latents``: None
        inputs = self.get_dummy_inputs(device)
        inputs["strength"] = 0.9
        out_0 = sd_pipe(**inputs).images

        # image latents + mask latents
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe.image_processor.preprocess(inputs["image"]).to(sd_pipe.device)
        mask = sd_pipe.mask_processor.preprocess(inputs["mask_image"]).to(sd_pipe.device)
        masked_image = image * (mask < 0.5)

        generator = torch.Generator(device=device).manual_seed(0)
        image_latents = (
            sd_pipe.vae.encode(image).latent_dist.sample(generator=generator) * sd_pipe.vae.config.scaling_factor
        )
        torch.randn((1, 4, 32, 32), generator=generator)
        mask_latents = (
            sd_pipe.vae.encode(masked_image).latent_dist.sample(generator=generator)
            * sd_pipe.vae.config.scaling_factor
        )
        inputs["image"] = image_latents
        inputs["masked_image_latents"] = mask_latents
        inputs["mask_image"] = mask
        inputs["strength"] = 0.9
        generator = torch.Generator(device=device).manual_seed(0)
        torch.randn((1, 4, 32, 32), generator=generator)
        inputs["generator"] = generator
        out_1 = sd_pipe(**inputs).images
        assert np.abs(out_0 - out_1).max() < 1e-2

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
class StableDiffusionSimpleInpaintPipelineFastTests(StableDiffusionInpaintPipelineFastTests):
    pipeline_class = StableDiffusionInpaintPipeline
    params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
    image_params = frozenset([])
    # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess

    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"),
            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,
        }
        return components

    def test_stable_diffusion_inpaint(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**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, 64, 64, 3)
        expected_slice = np.array([0.4925, 0.4967, 0.4100, 0.5234, 0.5322, 0.4532, 0.5805, 0.5877, 0.4151])

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

    @unittest.skip("skipped here because area stays unchanged due to mask")
    def test_stable_diffusion_inpaint_lora(self):
        ...


369
@slow
370
@require_torch_gpu
371
372
373
374
class StableDiffusionInpaintPipelineSlowTests(unittest.TestCase):
    def setUp(self):
        super().setUp()

375
376
377
378
379
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

380
381
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
382
        init_image = load_image(
383
384
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_image.png"
385
386
        )
        mask_image = load_image(
387
388
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_mask.png"
389
        )
390
391
392
393
394
395
396
397
398
399
        inputs = {
            "prompt": "Face of a yellow cat, high resolution, sitting on a park bench",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs
400

401
402
403
404
    def test_stable_diffusion_inpaint_ddim(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
405
406
407
408
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

409
410
411
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
412

413
        assert image.shape == (1, 512, 512, 3)
414
415
        expected_slice = np.array([0.0427, 0.0460, 0.0483, 0.0460, 0.0584, 0.0521, 0.1549, 0.1695, 0.1794])

416
        assert np.abs(expected_slice - image_slice).max() < 6e-4
417
418
419

    def test_stable_diffusion_inpaint_fp16(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
420
            "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16, safety_checker=None
421
        )
422
        pipe.unet.set_default_attn_processor()
423
424
425
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()
426

427
428
429
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
430

431
        assert image.shape == (1, 512, 512, 3)
432
        expected_slice = np.array([0.1509, 0.1245, 0.1672, 0.1655, 0.1519, 0.1226, 0.1462, 0.1567, 0.2451])
433
        assert np.abs(expected_slice - image_slice).max() < 5e-2
434

435
    def test_stable_diffusion_inpaint_pndm(self):
436
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
437
            "runwayml/stable-diffusion-inpainting", safety_checker=None
438
        )
439
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
440
441
442
443
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

444
445
446
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
447

448
        assert image.shape == (1, 512, 512, 3)
449
450
        expected_slice = np.array([0.0425, 0.0273, 0.0344, 0.1694, 0.1727, 0.1812, 0.3256, 0.3311, 0.3272])

451
        assert np.abs(expected_slice - image_slice).max() < 5e-3
452

453
454
455
    def test_stable_diffusion_inpaint_k_lms(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
456
        )
457
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
458
459
460
461
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

462
463
464
465
466
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
467
468
        expected_slice = np.array([0.9314, 0.7575, 0.9432, 0.8885, 0.9028, 0.7298, 0.9811, 0.9667, 0.7633])

469
        assert np.abs(expected_slice - image_slice).max() < 6e-3
470

471
472
473
474
475
476
    def test_stable_diffusion_inpaint_with_sequential_cpu_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe = StableDiffusionInpaintPipeline.from_pretrained(
477
            "runwayml/stable-diffusion-inpainting", safety_checker=None, torch_dtype=torch.float16
478
        )
479
480
481
482
483
484
485
486
487
488
489
490
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = 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

491
    @require_torch_2
492
    def test_inpaint_compile(self):
493
494
495
496
497
498
499
        seed = 0
        inputs = self.get_inputs(torch_device, seed=seed)
        # Can't pickle a Generator object
        del inputs["generator"]
        inputs["torch_device"] = torch_device
        inputs["seed"] = seed
        run_test_in_subprocess(test_case=self, target_func=_test_inpaint_compile, inputs=inputs)
500

501
    def test_stable_diffusion_inpaint_pil_input_resolution_test(self):
Patrick von Platen's avatar
Patrick von Platen committed
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", 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(torch_device)
        # change input image to a random size (one that would cause a tensor mismatch error)
        inputs["image"] = inputs["image"].resize((127, 127))
        inputs["mask_image"] = inputs["mask_image"].resize((127, 127))
        inputs["height"] = 128
        inputs["width"] = 128
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, inputs["height"], inputs["width"], 3)
519

520
521
522
523
524
    def test_stable_diffusion_inpaint_strength_test(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
525
        pipe.unet.set_default_attn_processor()
526
527
528
529
530
531
532
533
534
535
536
537
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        # change input strength
        inputs["strength"] = 0.75
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, 512, 512, 3)

        image_slice = image[0, 253:256, 253:256, -1].flatten()
538
539
        expected_slice = np.array([0.2728, 0.2803, 0.2665, 0.2511, 0.2774, 0.2586, 0.2391, 0.2392, 0.2582])
        assert np.abs(expected_slice - image_slice).max() < 1e-3
540

541
542
    def test_stable_diffusion_simple_inpaint_ddim(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", safety_checker=None)
543
        pipe.unet.set_default_attn_processor()
544
545
546
547
548
549
550
551
552
553
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

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

        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
554
555
        expected_slice = np.array([0.3757, 0.3875, 0.4445, 0.4353, 0.3780, 0.4513, 0.3965, 0.3984, 0.4362])
        assert np.abs(expected_slice - image_slice).max() < 1e-3
556

557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
    def test_download_local(self):
        filename = hf_hub_download("runwayml/stable-diffusion-inpainting", filename="sd-v1-5-inpainting.ckpt")

        pipe = StableDiffusionInpaintPipeline.from_single_file(filename, torch_dtype=torch.float16)
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 1
        image_out = pipe(**inputs).images[0]

        assert image_out.shape == (512, 512, 3)

    def test_download_ckpt_diff_format_is_same(self):
        ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-inpainting/blob/main/sd-v1-5-inpainting.ckpt"

        pipe = StableDiffusionInpaintPipeline.from_single_file(ckpt_path)
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.unet.set_attn_processor(AttnProcessor())
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 5
        image_ckpt = pipe(**inputs).images[0]

        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.unet.set_attn_processor(AttnProcessor())
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 5
        image = pipe(**inputs).images[0]

        assert np.max(np.abs(image - image_ckpt)) < 1e-4

593

Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
@slow
@require_torch_gpu
class StableDiffusionInpaintPipelineAsymmetricAutoencoderKLSlowTests(unittest.TestCase):
    def setUp(self):
        super().setUp()

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

    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
        init_image = load_image(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_image.png"
        )
        mask_image = load_image(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_mask.png"
        )
        inputs = {
            "prompt": "Face of a yellow cat, high resolution, sitting on a park bench",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs

    def test_stable_diffusion_inpaint_ddim(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.vae = vae
632
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
633
634
635
636
637
638
639
640
641
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
642
        expected_slice = np.array([0.0522, 0.0604, 0.0596, 0.0449, 0.0493, 0.0427, 0.1186, 0.1289, 0.1442])
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
643

644
        assert np.abs(expected_slice - image_slice).max() < 1e-3
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
645
646
647
648
649
650
651
652

    def test_stable_diffusion_inpaint_fp16(self):
        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5", torch_dtype=torch.float16
        )
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16, safety_checker=None
        )
653
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
        pipe.vae = vae
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.1343, 0.1406, 0.1440, 0.1504, 0.1729, 0.0989, 0.1807, 0.2822, 0.1179])

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

    def test_stable_diffusion_inpaint_pndm(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
673
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
674
675
676
677
678
679
680
681
682
683
684
        pipe.vae = vae
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
685
        expected_slice = np.array([0.0966, 0.1083, 0.1148, 0.1422, 0.1318, 0.1197, 0.3702, 0.3537, 0.3288])
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
686
687
688
689
690
691
692
693

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

    def test_stable_diffusion_inpaint_k_lms(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
694
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
695
696
697
698
699
700
701
702
703
704
        pipe.vae = vae
        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(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
        assert image.shape == (1, 512, 512, 3)
705
        expected_slice = np.array([0.8931, 0.8683, 0.8965, 0.8501, 0.8592, 0.9118, 0.8734, 0.7463, 0.8990])
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
        assert np.abs(expected_slice - image_slice).max() < 6e-3

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

        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5", torch_dtype=torch.float16
        )
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None, torch_dtype=torch.float16
        )
        pipe.vae = vae
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.45 GB is allocated
        assert mem_bytes < 2.45 * 10**9

    @require_torch_2
    def test_inpaint_compile(self):
        pass

    def test_stable_diffusion_inpaint_pil_input_resolution_test(self):
        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5",
        )
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.vae = vae
        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(torch_device)
        # change input image to a random size (one that would cause a tensor mismatch error)
        inputs["image"] = inputs["image"].resize((127, 127))
        inputs["mask_image"] = inputs["mask_image"].resize((127, 127))
        inputs["height"] = 128
        inputs["width"] = 128
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, inputs["height"], inputs["width"], 3)

    def test_stable_diffusion_inpaint_strength_test(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
764
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
        pipe.vae = vae
        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(torch_device)
        # change input strength
        inputs["strength"] = 0.75
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, 512, 512, 3)

        image_slice = image[0, 253:256, 253:256, -1].flatten()
        expected_slice = np.array([0.2458, 0.2576, 0.3124, 0.2679, 0.2669, 0.2796, 0.2872, 0.2975, 0.2661])
        assert np.abs(expected_slice - image_slice).max() < 3e-3

    def test_stable_diffusion_simple_inpaint_ddim(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", safety_checker=None)
        pipe.vae = vae
786
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
787
788
789
790
791
792
793
794
795
796
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

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

        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
797
798
        expected_slice = np.array([0.3296, 0.4041, 0.4097, 0.4145, 0.4342, 0.4152, 0.4927, 0.4931, 0.4430])
        assert np.abs(expected_slice - image_slice).max() < 1e-3
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820

    def test_download_local(self):
        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5", torch_dtype=torch.float16
        )
        filename = hf_hub_download("runwayml/stable-diffusion-inpainting", filename="sd-v1-5-inpainting.ckpt")

        pipe = StableDiffusionInpaintPipeline.from_single_file(filename, torch_dtype=torch.float16)
        pipe.vae = vae
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 1
        image_out = pipe(**inputs).images[0]

        assert image_out.shape == (512, 512, 3)

    def test_download_ckpt_diff_format_is_same(self):
        pass


821
822
823
824
825
826
827
@nightly
@require_torch_gpu
class StableDiffusionInpaintPipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()
828

829
830
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
831
        init_image = load_image(
832
833
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_image.png"
834
835
        )
        mask_image = load_image(
836
837
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_mask.png"
838
        )
839
840
841
842
843
844
845
846
847
848
        inputs = {
            "prompt": "Face of a yellow cat, high resolution, sitting on a park bench",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 50,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs
849

850
851
852
853
    def test_inpaint_ddim(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
854

855
856
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]
857

858
859
860
861
862
863
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
864

865
866
867
868
869
870
871
872
    def test_inpaint_pndm(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.scheduler = PNDMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]
873

874
875
876
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_pndm.npy"
877
        )
878
879
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
880

881
882
883
884
885
    def test_inpaint_lms(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
886

887
888
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]
889

890
891
892
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_lms.npy"
893
        )
894
895
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
896

897
898
899
900
901
    def test_inpaint_dpm(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.scheduler = DPMSolverMultistepScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
902

903
904
905
        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 30
        image = sd_pipe(**inputs).images[0]
906

907
908
909
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_dpm_multi.npy"
910
        )
911
912
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
913

Patrick von Platen's avatar
Patrick von Platen committed
914

915
916
class StableDiffusionInpaintingPrepareMaskAndMaskedImageTests(unittest.TestCase):
    def test_pil_inputs(self):
917
918
        height, width = 32, 32
        im = np.random.randint(0, 255, (height, width, 3), dtype=np.uint8)
919
        im = Image.fromarray(im)
920
        mask = np.random.randint(0, 255, (height, width), dtype=np.uint8) > 127.5
921
922
        mask = Image.fromarray((mask * 255).astype(np.uint8))

923
        t_mask, t_masked, t_image = prepare_mask_and_masked_image(im, mask, height, width, return_image=True)
924
925
926

        self.assertTrue(isinstance(t_mask, torch.Tensor))
        self.assertTrue(isinstance(t_masked, torch.Tensor))
927
        self.assertTrue(isinstance(t_image, torch.Tensor))
928
929
930

        self.assertEqual(t_mask.ndim, 4)
        self.assertEqual(t_masked.ndim, 4)
931
        self.assertEqual(t_image.ndim, 4)
932

933
934
        self.assertEqual(t_mask.shape, (1, 1, height, width))
        self.assertEqual(t_masked.shape, (1, 3, height, width))
935
        self.assertEqual(t_image.shape, (1, 3, height, width))
936
937
938

        self.assertTrue(t_mask.dtype == torch.float32)
        self.assertTrue(t_masked.dtype == torch.float32)
939
        self.assertTrue(t_image.dtype == torch.float32)
940
941
942
943
944

        self.assertTrue(t_mask.min() >= 0.0)
        self.assertTrue(t_mask.max() <= 1.0)
        self.assertTrue(t_masked.min() >= -1.0)
        self.assertTrue(t_masked.min() <= 1.0)
945
946
        self.assertTrue(t_image.min() >= -1.0)
        self.assertTrue(t_image.min() >= -1.0)
947
948
949
950

        self.assertTrue(t_mask.sum() > 0.0)

    def test_np_inputs(self):
951
952
953
        height, width = 32, 32

        im_np = np.random.randint(0, 255, (height, width, 3), dtype=np.uint8)
954
        im_pil = Image.fromarray(im_np)
Patrick von Platen's avatar
Patrick von Platen committed
955
956
957
958
959
960
961
962
963
964
965
966
        mask_np = (
            np.random.randint(
                0,
                255,
                (
                    height,
                    width,
                ),
                dtype=np.uint8,
            )
            > 127.5
        )
967
968
        mask_pil = Image.fromarray((mask_np * 255).astype(np.uint8))

969
970
971
972
973
974
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
        )
        t_mask_pil, t_masked_pil, t_image_pil = prepare_mask_and_masked_image(
            im_pil, mask_pil, height, width, return_image=True
        )
975
976
977

        self.assertTrue((t_mask_np == t_mask_pil).all())
        self.assertTrue((t_masked_np == t_masked_pil).all())
978
        self.assertTrue((t_image_np == t_image_pil).all())
979
980

    def test_torch_3D_2D_inputs(self):
981
982
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
        im_tensor = torch.randint(
            0,
            255,
            (
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1005
1006
1007
        im_np = im_tensor.numpy().transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()

1008
1009
1010
1011
1012
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1013
        )
1014
1015
1016

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1017
        self.assertTrue((t_image_tensor == t_image_np).all())
1018
1019

    def test_torch_3D_3D_inputs(self):
1020
1021
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
        im_tensor = torch.randint(
            0,
            255,
            (
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1045
1046
1047
        im_np = im_tensor.numpy().transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()[0]

1048
1049
1050
1051
1052
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1053
        )
1054
1055
1056

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1057
        self.assertTrue((t_image_tensor == t_image_np).all())
1058
1059

    def test_torch_4D_2D_inputs(self):
1060
1061
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
        im_tensor = torch.randint(
            0,
            255,
            (
                1,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1085
1086
1087
        im_np = im_tensor.numpy()[0].transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()

1088
1089
1090
1091
1092
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1093
        )
1094
1095
1096

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1097
        self.assertTrue((t_image_tensor == t_image_np).all())
1098
1099

    def test_torch_4D_3D_inputs(self):
1100
1101
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
        im_tensor = torch.randint(
            0,
            255,
            (
                1,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1126
1127
1128
        im_np = im_tensor.numpy()[0].transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()[0]

1129
1130
1131
1132
1133
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1134
        )
1135
1136
1137

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1138
        self.assertTrue((t_image_tensor == t_image_np).all())
1139
1140

    def test_torch_4D_4D_inputs(self):
1141
1142
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
        im_tensor = torch.randint(
            0,
            255,
            (
                1,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    1,
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1168
1169
1170
        im_np = im_tensor.numpy()[0].transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()[0][0]

1171
1172
1173
1174
1175
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1176
        )
1177
1178
1179

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1180
        self.assertTrue((t_image_tensor == t_image_np).all())
1181
1182

    def test_torch_batch_4D_3D(self):
1183
1184
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
        im_tensor = torch.randint(
            0,
            255,
            (
                2,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    2,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1209
1210
1211
1212

        im_nps = [im.numpy().transpose(1, 2, 0) for im in im_tensor]
        mask_nps = [mask.numpy() for mask in mask_tensor]

1213
1214
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1215
        )
1216
        nps = [prepare_mask_and_masked_image(i, m, height, width, return_image=True) for i, m in zip(im_nps, mask_nps)]
1217
1218
        t_mask_np = torch.cat([n[0] for n in nps])
        t_masked_np = torch.cat([n[1] for n in nps])
1219
        t_image_np = torch.cat([n[2] for n in nps])
1220
1221
1222

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1223
        self.assertTrue((t_image_tensor == t_image_np).all())
1224
1225

    def test_torch_batch_4D_4D(self):
1226
1227
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
        im_tensor = torch.randint(
            0,
            255,
            (
                2,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    2,
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1253
1254
1255
1256

        im_nps = [im.numpy().transpose(1, 2, 0) for im in im_tensor]
        mask_nps = [mask.numpy()[0] for mask in mask_tensor]

1257
1258
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1259
        )
1260
        nps = [prepare_mask_and_masked_image(i, m, height, width, return_image=True) for i, m in zip(im_nps, mask_nps)]
1261
1262
        t_mask_np = torch.cat([n[0] for n in nps])
        t_masked_np = torch.cat([n[1] for n in nps])
1263
        t_image_np = torch.cat([n[2] for n in nps])
1264
1265
1266

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1267
        self.assertTrue((t_image_tensor == t_image_np).all())
1268
1269

    def test_shape_mismatch(self):
1270
1271
        height, width = 32, 32

1272
1273
        # test height and width
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1274
1275
1276
1277
1278
1279
1280
1281
1282
            prepare_mask_and_masked_image(
                torch.randn(
                    3,
                    height,
                    width,
                ),
                torch.randn(64, 64),
                height,
                width,
1283
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1284
            )
1285
1286
        # test batch dim
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
            prepare_mask_and_masked_image(
                torch.randn(
                    2,
                    3,
                    height,
                    width,
                ),
                torch.randn(4, 64, 64),
                height,
                width,
1297
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1298
            )
1299
1300
        # test batch dim
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
            prepare_mask_and_masked_image(
                torch.randn(
                    2,
                    3,
                    height,
                    width,
                ),
                torch.randn(4, 1, 64, 64),
                height,
                width,
1311
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1312
            )
1313
1314

    def test_type_mismatch(self):
1315
1316
        height, width = 32, 32

1317
1318
        # test tensors-only
        with self.assertRaises(TypeError):
Patrick von Platen's avatar
Patrick von Platen committed
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ),
                torch.rand(
                    3,
                    height,
                    width,
                ).numpy(),
                height,
                width,
1332
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1333
            )
1334
1335
        # test tensors-only
        with self.assertRaises(TypeError):
Patrick von Platen's avatar
Patrick von Platen committed
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ).numpy(),
                torch.rand(
                    3,
                    height,
                    width,
                ),
                height,
                width,
1349
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1350
            )
1351
1352

    def test_channels_first(self):
1353
1354
        height, width = 32, 32

1355
1356
        # test channels first for 3D tensors
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1357
1358
1359
1360
1361
1362
1363
1364
1365
            prepare_mask_and_masked_image(
                torch.rand(height, width, 3),
                torch.rand(
                    3,
                    height,
                    width,
                ),
                height,
                width,
1366
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1367
            )
1368
1369

    def test_tensor_range(self):
1370
1371
        height, width = 32, 32

1372
1373
        # test im <= 1
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
            prepare_mask_and_masked_image(
                torch.ones(
                    3,
                    height,
                    width,
                )
                * 2,
                torch.rand(
                    height,
                    width,
                ),
                height,
                width,
1387
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1388
            )
1389
1390
        # test im >= -1
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
            prepare_mask_and_masked_image(
                torch.ones(
                    3,
                    height,
                    width,
                )
                * (-2),
                torch.rand(
                    height,
                    width,
                ),
                height,
                width,
1404
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1405
            )
1406
1407
        # test mask <= 1
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ),
                torch.ones(
                    height,
                    width,
                )
                * 2,
                height,
                width,
1421
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1422
            )
1423
1424
        # test mask >= 0
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ),
                torch.ones(
                    height,
                    width,
                )
                * -1,
                height,
                width,
1438
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1439
            )