test_transforms_v2_consistency.py 51.1 KB
Newer Older
1
import enum
2
3
import importlib.machinery
import importlib.util
4
import inspect
5
import random
6
import re
7
from collections import defaultdict
8
from pathlib import Path
9

10
import numpy as np
11
import PIL.Image
12
import pytest
13
14

import torch
15
import torchvision.transforms.v2 as v2_transforms
16
from common_utils import (
17
    ArgsKwargs,
18
    assert_close,
19
20
21
22
23
    assert_equal,
    make_bounding_box,
    make_detection_mask,
    make_image,
    make_images,
24
    make_segmentation_mask,
25
)
26
from torch import nn
27
from torchvision import datapoints, transforms as legacy_transforms
28
from torchvision._utils import sequence_to_str
29

30
from torchvision.transforms import functional as legacy_F
31
32
33
from torchvision.transforms.v2 import functional as prototype_F
from torchvision.transforms.v2.functional import to_image_pil
from torchvision.transforms.v2.utils import query_spatial_size
34

35
DEFAULT_MAKE_IMAGES_KWARGS = dict(color_spaces=["RGB"], extra_dims=[(4,)])
36
37


38
39
40
41
42
43
44
45
46
class NotScriptableArgsKwargs(ArgsKwargs):
    """
    This class is used to mark parameters that render the transform non-scriptable. They still work in eager mode and
    thus will be tested there, but will be skipped by the JIT tests.
    """

    pass


47
48
class ConsistencyConfig:
    def __init__(
49
50
51
        self,
        prototype_cls,
        legacy_cls,
52
53
        # If no args_kwargs is passed, only the signature will be checked
        args_kwargs=(),
54
55
56
        make_images_kwargs=None,
        supports_pil=True,
        removed_params=(),
57
        closeness_kwargs=None,
58
59
60
    ):
        self.prototype_cls = prototype_cls
        self.legacy_cls = legacy_cls
61
        self.args_kwargs = args_kwargs
62
63
        self.make_images_kwargs = make_images_kwargs or DEFAULT_MAKE_IMAGES_KWARGS
        self.supports_pil = supports_pil
64
        self.removed_params = removed_params
65
        self.closeness_kwargs = closeness_kwargs or dict(rtol=0, atol=0)
66
67


68
69
70
71
# These are here since both the prototype and legacy transform need to be constructed with the same random parameters
LINEAR_TRANSFORMATION_MEAN = torch.rand(36)
LINEAR_TRANSFORMATION_MATRIX = torch.rand([LINEAR_TRANSFORMATION_MEAN.numel()] * 2)

72
73
CONSISTENCY_CONFIGS = [
    ConsistencyConfig(
74
        v2_transforms.Normalize,
75
76
77
78
79
80
81
82
        legacy_transforms.Normalize,
        [
            ArgsKwargs(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
        ],
        supports_pil=False,
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, dtypes=[torch.float]),
    ),
    ConsistencyConfig(
83
        v2_transforms.Resize,
84
85
        legacy_transforms.Resize,
        [
86
            NotScriptableArgsKwargs(32),
87
            ArgsKwargs([32]),
88
            ArgsKwargs((32, 29)),
89
90
            ArgsKwargs((31, 28), interpolation=v2_transforms.InterpolationMode.NEAREST),
            ArgsKwargs((33, 26), interpolation=v2_transforms.InterpolationMode.BICUBIC),
91
92
93
            ArgsKwargs((30, 27), interpolation=PIL.Image.NEAREST),
            ArgsKwargs((35, 29), interpolation=PIL.Image.BILINEAR),
            ArgsKwargs((34, 25), interpolation=PIL.Image.BICUBIC),
94
95
96
97
            NotScriptableArgsKwargs(31, max_size=32),
            ArgsKwargs([31], max_size=32),
            NotScriptableArgsKwargs(30, max_size=100),
            ArgsKwargs([31], max_size=32),
98
99
            ArgsKwargs((29, 32), antialias=False),
            ArgsKwargs((28, 31), antialias=True),
100
        ],
101
102
        # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
        closeness_kwargs=dict(rtol=0, atol=1),
103
104
    ),
    ConsistencyConfig(
105
        v2_transforms.CenterCrop,
106
107
108
109
110
111
        legacy_transforms.CenterCrop,
        [
            ArgsKwargs(18),
            ArgsKwargs((18, 13)),
        ],
    ),
112
    ConsistencyConfig(
113
        v2_transforms.FiveCrop,
114
115
116
117
118
119
120
121
        legacy_transforms.FiveCrop,
        [
            ArgsKwargs(18),
            ArgsKwargs((18, 13)),
        ],
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, sizes=[(20, 19)]),
    ),
    ConsistencyConfig(
122
        v2_transforms.TenCrop,
123
124
125
126
        legacy_transforms.TenCrop,
        [
            ArgsKwargs(18),
            ArgsKwargs((18, 13)),
127
            ArgsKwargs(18, vertical_flip=True),
128
129
130
131
        ],
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, sizes=[(20, 19)]),
    ),
    ConsistencyConfig(
132
        v2_transforms.Pad,
133
134
        legacy_transforms.Pad,
        [
135
            NotScriptableArgsKwargs(3),
136
137
138
            ArgsKwargs([3]),
            ArgsKwargs([2, 3]),
            ArgsKwargs([3, 2, 1, 4]),
139
140
141
142
143
            NotScriptableArgsKwargs(5, fill=1, padding_mode="constant"),
            ArgsKwargs([5], fill=1, padding_mode="constant"),
            NotScriptableArgsKwargs(5, padding_mode="edge"),
            NotScriptableArgsKwargs(5, padding_mode="reflect"),
            NotScriptableArgsKwargs(5, padding_mode="symmetric"),
144
145
        ],
    ),
146
147
    *[
        ConsistencyConfig(
148
            v2_transforms.LinearTransformation,
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
            legacy_transforms.LinearTransformation,
            [
                ArgsKwargs(LINEAR_TRANSFORMATION_MATRIX.to(matrix_dtype), LINEAR_TRANSFORMATION_MEAN.to(matrix_dtype)),
            ],
            # Make sure that the product of the height, width and number of channels matches the number of elements in
            # `LINEAR_TRANSFORMATION_MEAN`. For example 2 * 6 * 3 == 4 * 3 * 3 == 36.
            make_images_kwargs=dict(
                DEFAULT_MAKE_IMAGES_KWARGS, sizes=[(2, 6), (4, 3)], color_spaces=["RGB"], dtypes=[image_dtype]
            ),
            supports_pil=False,
        )
        for matrix_dtype, image_dtype in [
            (torch.float32, torch.float32),
            (torch.float64, torch.float64),
            (torch.float32, torch.uint8),
            (torch.float64, torch.float32),
            (torch.float32, torch.float64),
        ]
    ],
168
    ConsistencyConfig(
169
        v2_transforms.Grayscale,
170
171
172
173
174
        legacy_transforms.Grayscale,
        [
            ArgsKwargs(num_output_channels=1),
            ArgsKwargs(num_output_channels=3),
        ],
175
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, color_spaces=["RGB", "GRAY"]),
176
177
        # Use default tolerances of `torch.testing.assert_close`
        closeness_kwargs=dict(rtol=None, atol=None),
178
    ),
179
    ConsistencyConfig(
180
        v2_transforms.ConvertDtype,
181
182
183
184
185
186
187
188
189
        legacy_transforms.ConvertImageDtype,
        [
            ArgsKwargs(torch.float16),
            ArgsKwargs(torch.bfloat16),
            ArgsKwargs(torch.float32),
            ArgsKwargs(torch.float64),
            ArgsKwargs(torch.uint8),
        ],
        supports_pil=False,
190
191
        # Use default tolerances of `torch.testing.assert_close`
        closeness_kwargs=dict(rtol=None, atol=None),
192
193
    ),
    ConsistencyConfig(
194
        v2_transforms.ToPILImage,
195
        legacy_transforms.ToPILImage,
196
        [NotScriptableArgsKwargs()],
197
198
        make_images_kwargs=dict(
            color_spaces=[
199
200
201
202
                "GRAY",
                "GRAY_ALPHA",
                "RGB",
                "RGBA",
203
204
205
206
207
208
            ],
            extra_dims=[()],
        ),
        supports_pil=False,
    ),
    ConsistencyConfig(
209
        v2_transforms.Lambda,
210
211
        legacy_transforms.Lambda,
        [
212
            NotScriptableArgsKwargs(lambda image: image / 2),
213
214
215
216
217
        ],
        # Technically, this also supports PIL, but it is overkill to write a function here that supports tensor and PIL
        # images given that the transform does nothing but call it anyway.
        supports_pil=False,
    ),
218
    ConsistencyConfig(
219
        v2_transforms.RandomHorizontalFlip,
220
221
222
223
224
225
226
        legacy_transforms.RandomHorizontalFlip,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
        ],
    ),
    ConsistencyConfig(
227
        v2_transforms.RandomVerticalFlip,
228
229
230
231
232
233
234
        legacy_transforms.RandomVerticalFlip,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
        ],
    ),
    ConsistencyConfig(
235
        v2_transforms.RandomEqualize,
236
237
238
239
240
241
242
243
        legacy_transforms.RandomEqualize,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
        ],
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, dtypes=[torch.uint8]),
    ),
    ConsistencyConfig(
244
        v2_transforms.RandomInvert,
245
246
247
248
249
250
251
        legacy_transforms.RandomInvert,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
        ],
    ),
    ConsistencyConfig(
252
        v2_transforms.RandomPosterize,
253
254
255
256
257
258
259
260
261
        legacy_transforms.RandomPosterize,
        [
            ArgsKwargs(p=0, bits=5),
            ArgsKwargs(p=1, bits=1),
            ArgsKwargs(p=1, bits=3),
        ],
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, dtypes=[torch.uint8]),
    ),
    ConsistencyConfig(
262
        v2_transforms.RandomSolarize,
263
264
265
266
267
268
269
        legacy_transforms.RandomSolarize,
        [
            ArgsKwargs(p=0, threshold=0.5),
            ArgsKwargs(p=1, threshold=0.3),
            ArgsKwargs(p=1, threshold=0.99),
        ],
    ),
270
271
    *[
        ConsistencyConfig(
272
            v2_transforms.RandomAutocontrast,
273
274
275
276
277
278
279
280
281
282
            legacy_transforms.RandomAutocontrast,
            [
                ArgsKwargs(p=0),
                ArgsKwargs(p=1),
            ],
            make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, dtypes=[dt]),
            closeness_kwargs=ckw,
        )
        for dt, ckw in [(torch.uint8, dict(atol=1, rtol=0)), (torch.float32, dict(rtol=None, atol=None))]
    ],
283
    ConsistencyConfig(
284
        v2_transforms.RandomAdjustSharpness,
285
286
287
        legacy_transforms.RandomAdjustSharpness,
        [
            ArgsKwargs(p=0, sharpness_factor=0.5),
288
            ArgsKwargs(p=1, sharpness_factor=0.2),
289
290
            ArgsKwargs(p=1, sharpness_factor=0.99),
        ],
291
        closeness_kwargs={"atol": 1e-6, "rtol": 1e-6},
292
293
    ),
    ConsistencyConfig(
294
        v2_transforms.RandomGrayscale,
295
296
297
298
299
        legacy_transforms.RandomGrayscale,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
        ],
300
301
302
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, color_spaces=["RGB", "GRAY"]),
        # Use default tolerances of `torch.testing.assert_close`
        closeness_kwargs=dict(rtol=None, atol=None),
303
304
    ),
    ConsistencyConfig(
305
        v2_transforms.RandomResizedCrop,
306
307
308
309
310
        legacy_transforms.RandomResizedCrop,
        [
            ArgsKwargs(16),
            ArgsKwargs(17, scale=(0.3, 0.7)),
            ArgsKwargs(25, ratio=(0.5, 1.5)),
311
312
            ArgsKwargs((31, 28), interpolation=v2_transforms.InterpolationMode.NEAREST),
            ArgsKwargs((33, 26), interpolation=v2_transforms.InterpolationMode.BICUBIC),
313
314
            ArgsKwargs((31, 28), interpolation=PIL.Image.NEAREST),
            ArgsKwargs((33, 26), interpolation=PIL.Image.BICUBIC),
315
316
317
            ArgsKwargs((29, 32), antialias=False),
            ArgsKwargs((28, 31), antialias=True),
        ],
318
319
        # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
        closeness_kwargs=dict(rtol=0, atol=1),
320
321
    ),
    ConsistencyConfig(
322
        v2_transforms.RandomErasing,
323
324
325
326
327
328
329
330
331
332
333
334
335
        legacy_transforms.RandomErasing,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
            ArgsKwargs(p=1, scale=(0.3, 0.7)),
            ArgsKwargs(p=1, ratio=(0.5, 1.5)),
            ArgsKwargs(p=1, value=1),
            ArgsKwargs(p=1, value=(1, 2, 3)),
            ArgsKwargs(p=1, value="random"),
        ],
        supports_pil=False,
    ),
    ConsistencyConfig(
336
        v2_transforms.ColorJitter,
337
338
339
340
341
342
343
344
345
346
347
        legacy_transforms.ColorJitter,
        [
            ArgsKwargs(),
            ArgsKwargs(brightness=0.1),
            ArgsKwargs(brightness=(0.2, 0.3)),
            ArgsKwargs(contrast=0.4),
            ArgsKwargs(contrast=(0.5, 0.6)),
            ArgsKwargs(saturation=0.7),
            ArgsKwargs(saturation=(0.8, 0.9)),
            ArgsKwargs(hue=0.3),
            ArgsKwargs(hue=(-0.1, 0.2)),
348
            ArgsKwargs(brightness=0.1, contrast=0.4, saturation=0.5, hue=0.3),
349
        ],
350
        closeness_kwargs={"atol": 1e-5, "rtol": 1e-5},
351
    ),
352
353
    *[
        ConsistencyConfig(
354
            v2_transforms.ElasticTransform,
355
356
357
358
359
360
361
            legacy_transforms.ElasticTransform,
            [
                ArgsKwargs(),
                ArgsKwargs(alpha=20.0),
                ArgsKwargs(alpha=(15.3, 27.2)),
                ArgsKwargs(sigma=3.0),
                ArgsKwargs(sigma=(2.5, 3.9)),
362
363
                ArgsKwargs(interpolation=v2_transforms.InterpolationMode.NEAREST),
                ArgsKwargs(interpolation=v2_transforms.InterpolationMode.BICUBIC),
364
365
                ArgsKwargs(interpolation=PIL.Image.NEAREST),
                ArgsKwargs(interpolation=PIL.Image.BICUBIC),
366
367
368
369
370
371
372
373
374
375
                ArgsKwargs(fill=1),
            ],
            # ElasticTransform needs larger images to avoid the needed internal padding being larger than the actual image
            make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, sizes=[(163, 163), (72, 333), (313, 95)], dtypes=[dt]),
            # We updated gaussian blur kernel generation with a faster and numerically more stable version
            # This brings float32 accumulation visible in elastic transform -> we need to relax consistency tolerance
            closeness_kwargs=ckw,
        )
        for dt, ckw in [(torch.uint8, {"rtol": 1e-1, "atol": 1}), (torch.float32, {"rtol": 1e-2, "atol": 1e-3})]
    ],
376
    ConsistencyConfig(
377
        v2_transforms.GaussianBlur,
378
379
380
381
382
383
384
        legacy_transforms.GaussianBlur,
        [
            ArgsKwargs(kernel_size=3),
            ArgsKwargs(kernel_size=(1, 5)),
            ArgsKwargs(kernel_size=3, sigma=0.7),
            ArgsKwargs(kernel_size=5, sigma=(0.3, 1.4)),
        ],
385
        closeness_kwargs={"rtol": 1e-5, "atol": 1e-5},
386
387
    ),
    ConsistencyConfig(
388
        v2_transforms.RandomAffine,
389
390
391
392
393
394
395
396
397
398
        legacy_transforms.RandomAffine,
        [
            ArgsKwargs(degrees=30.0),
            ArgsKwargs(degrees=(-20.0, 10.0)),
            ArgsKwargs(degrees=0.0, translate=(0.4, 0.6)),
            ArgsKwargs(degrees=0.0, scale=(0.3, 0.8)),
            ArgsKwargs(degrees=0.0, shear=13),
            ArgsKwargs(degrees=0.0, shear=(8, 17)),
            ArgsKwargs(degrees=0.0, shear=(4, 5, 4, 13)),
            ArgsKwargs(degrees=(-20.0, 10.0), translate=(0.4, 0.6), scale=(0.3, 0.8), shear=(4, 5, 4, 13)),
399
            ArgsKwargs(degrees=30.0, interpolation=v2_transforms.InterpolationMode.NEAREST),
400
            ArgsKwargs(degrees=30.0, interpolation=PIL.Image.NEAREST),
401
402
403
404
            ArgsKwargs(degrees=30.0, fill=1),
            ArgsKwargs(degrees=30.0, fill=(2, 3, 4)),
            ArgsKwargs(degrees=30.0, center=(0, 0)),
        ],
405
        removed_params=["fillcolor", "resample"],
406
407
    ),
    ConsistencyConfig(
408
        v2_transforms.RandomCrop,
409
410
411
412
        legacy_transforms.RandomCrop,
        [
            ArgsKwargs(12),
            ArgsKwargs((15, 17)),
413
414
            NotScriptableArgsKwargs(11, padding=1),
            ArgsKwargs(11, padding=[1]),
415
416
417
418
            ArgsKwargs((8, 13), padding=(2, 3)),
            ArgsKwargs((14, 9), padding=(0, 2, 1, 0)),
            ArgsKwargs(36, pad_if_needed=True),
            ArgsKwargs((7, 8), fill=1),
419
            NotScriptableArgsKwargs(5, fill=(1, 2, 3)),
420
            ArgsKwargs(12),
421
            NotScriptableArgsKwargs(15, padding=2, padding_mode="edge"),
422
423
424
425
426
427
            ArgsKwargs(17, padding=(1, 0), padding_mode="reflect"),
            ArgsKwargs(8, padding=(3, 0, 0, 1), padding_mode="symmetric"),
        ],
        make_images_kwargs=dict(DEFAULT_MAKE_IMAGES_KWARGS, sizes=[(26, 26), (18, 33), (29, 22)]),
    ),
    ConsistencyConfig(
428
        v2_transforms.RandomPerspective,
429
430
431
432
433
        legacy_transforms.RandomPerspective,
        [
            ArgsKwargs(p=0),
            ArgsKwargs(p=1),
            ArgsKwargs(p=1, distortion_scale=0.3),
434
            ArgsKwargs(p=1, distortion_scale=0.2, interpolation=v2_transforms.InterpolationMode.NEAREST),
435
            ArgsKwargs(p=1, distortion_scale=0.2, interpolation=PIL.Image.NEAREST),
436
437
438
            ArgsKwargs(p=1, distortion_scale=0.1, fill=1),
            ArgsKwargs(p=1, distortion_scale=0.4, fill=(1, 2, 3)),
        ],
439
        closeness_kwargs={"atol": None, "rtol": None},
440
441
    ),
    ConsistencyConfig(
442
        v2_transforms.RandomRotation,
443
444
445
446
        legacy_transforms.RandomRotation,
        [
            ArgsKwargs(degrees=30.0),
            ArgsKwargs(degrees=(-20.0, 10.0)),
447
            ArgsKwargs(degrees=30.0, interpolation=v2_transforms.InterpolationMode.BILINEAR),
448
            ArgsKwargs(degrees=30.0, interpolation=PIL.Image.BILINEAR),
449
450
451
452
453
            ArgsKwargs(degrees=30.0, expand=True),
            ArgsKwargs(degrees=30.0, center=(0, 0)),
            ArgsKwargs(degrees=30.0, fill=1),
            ArgsKwargs(degrees=30.0, fill=(1, 2, 3)),
        ],
454
        removed_params=["resample"],
455
    ),
456
    ConsistencyConfig(
457
        v2_transforms.PILToTensor,
458
459
460
        legacy_transforms.PILToTensor,
    ),
    ConsistencyConfig(
461
        v2_transforms.ToTensor,
462
463
464
        legacy_transforms.ToTensor,
    ),
    ConsistencyConfig(
465
        v2_transforms.Compose,
466
467
468
        legacy_transforms.Compose,
    ),
    ConsistencyConfig(
469
        v2_transforms.RandomApply,
470
471
472
        legacy_transforms.RandomApply,
    ),
    ConsistencyConfig(
473
        v2_transforms.RandomChoice,
474
475
476
        legacy_transforms.RandomChoice,
    ),
    ConsistencyConfig(
477
        v2_transforms.RandomOrder,
478
479
480
        legacy_transforms.RandomOrder,
    ),
    ConsistencyConfig(
481
        v2_transforms.AugMix,
482
483
484
        legacy_transforms.AugMix,
    ),
    ConsistencyConfig(
485
        v2_transforms.AutoAugment,
486
487
488
        legacy_transforms.AutoAugment,
    ),
    ConsistencyConfig(
489
        v2_transforms.RandAugment,
490
491
492
        legacy_transforms.RandAugment,
    ),
    ConsistencyConfig(
493
        v2_transforms.TrivialAugmentWide,
494
495
        legacy_transforms.TrivialAugmentWide,
    ),
496
497
498
]


499
500
def test_automatic_coverage():
    available = {
501
502
        name
        for name, obj in legacy_transforms.__dict__.items()
503
        if not name.startswith("_") and isinstance(obj, type) and not issubclass(obj, enum.Enum)
504
505
    }

506
    checked = {config.legacy_cls.__name__ for config in CONSISTENCY_CONFIGS}
507

508
    missing = available - checked
509
510
511
512
513
514
515
    if missing:
        raise AssertionError(
            f"The prototype transformations {sequence_to_str(sorted(missing), separate_last='and ')} "
            f"are not checked for consistency although a legacy counterpart exists."
        )


516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
@pytest.mark.parametrize("config", CONSISTENCY_CONFIGS, ids=lambda config: config.legacy_cls.__name__)
def test_signature_consistency(config):
    legacy_params = dict(inspect.signature(config.legacy_cls).parameters)
    prototype_params = dict(inspect.signature(config.prototype_cls).parameters)

    for param in config.removed_params:
        legacy_params.pop(param, None)

    missing = legacy_params.keys() - prototype_params.keys()
    if missing:
        raise AssertionError(
            f"The prototype transform does not support the parameters "
            f"{sequence_to_str(sorted(missing), separate_last='and ')}, but the legacy transform does. "
            f"If that is intentional, e.g. pending deprecation, please add the parameters to the `removed_params` on "
            f"the `ConsistencyConfig`."
        )

    extra = prototype_params.keys() - legacy_params.keys()
534
535
536
537
538
539
    extra_without_default = {
        param
        for param in extra
        if prototype_params[param].default is inspect.Parameter.empty
        and prototype_params[param].kind not in {inspect.Parameter.VAR_POSITIONAL, inspect.Parameter.VAR_KEYWORD}
    }
540
541
    if extra_without_default:
        raise AssertionError(
542
543
544
            f"The prototype transform requires the parameters "
            f"{sequence_to_str(sorted(extra_without_default), separate_last='and ')}, but the legacy transform does "
            f"not. Please add a default value."
545
546
        )

547
548
549
550
551
552
    legacy_signature = list(legacy_params.keys())
    # Since we made sure that we don't have any extra parameters without default above, we clamp the prototype signature
    # to the same number of parameters as the legacy one
    prototype_signature = list(prototype_params.keys())[: len(legacy_signature)]

    assert prototype_signature == legacy_signature
553
554


555
556
557
def check_call_consistency(
    prototype_transform, legacy_transform, images=None, supports_pil=True, closeness_kwargs=None
):
558
559
    if images is None:
        images = make_images(**DEFAULT_MAKE_IMAGES_KWARGS)
560

561
562
    closeness_kwargs = closeness_kwargs or dict()

563
564
    for image in images:
        image_repr = f"[{tuple(image.shape)}, {str(image.dtype).rsplit('.')[-1]}]"
565
566
567

        image_tensor = torch.Tensor(image)
        try:
568
            torch.manual_seed(0)
569
            output_legacy_tensor = legacy_transform(image_tensor)
570
571
        except Exception as exc:
            raise pytest.UsageError(
572
                f"Transforming a tensor image {image_repr} failed in the legacy transform with the "
573
                f"error above. This means that you need to specify the parameters passed to `make_images` through the "
574
575
576
577
                "`make_images_kwargs` of the `ConsistencyConfig`."
            ) from exc

        try:
578
            torch.manual_seed(0)
579
            output_prototype_tensor = prototype_transform(image_tensor)
580
581
        except Exception as exc:
            raise AssertionError(
582
                f"Transforming a tensor image with shape {image_repr} failed in the prototype transform with "
583
584
                f"the error above. This means there is a consistency bug either in `_get_params` or in the "
                f"`is_simple_tensor` path in `_transform`."
585
586
            ) from exc

587
        assert_close(
588
589
590
            output_prototype_tensor,
            output_legacy_tensor,
            msg=lambda msg: f"Tensor image consistency check failed with: \n\n{msg}",
591
            **closeness_kwargs,
592
593
594
        )

        try:
595
            torch.manual_seed(0)
596
            output_prototype_image = prototype_transform(image)
597
598
        except Exception as exc:
            raise AssertionError(
599
                f"Transforming a image datapoint with shape {image_repr} failed in the prototype transform with "
600
                f"the error above. This means there is a consistency bug either in `_get_params` or in the "
601
                f"`datapoints.Image` path in `_transform`."
602
603
            ) from exc

604
        assert_close(
605
            output_prototype_image,
606
            output_prototype_tensor,
607
            msg=lambda msg: f"Output for datapoint and tensor images is not equal: \n\n{msg}",
608
            **closeness_kwargs,
609
610
        )

611
612
613
        if image.ndim == 3 and supports_pil:
            image_pil = to_image_pil(image)

614
            try:
615
                torch.manual_seed(0)
616
                output_legacy_pil = legacy_transform(image_pil)
617
618
            except Exception as exc:
                raise pytest.UsageError(
619
                    f"Transforming a PIL image with shape {image_repr} failed in the legacy transform with the "
620
621
622
623
624
                    f"error above. If this transform does not support PIL images, set `supports_pil=False` on the "
                    "`ConsistencyConfig`. "
                ) from exc

            try:
625
                torch.manual_seed(0)
626
                output_prototype_pil = prototype_transform(image_pil)
627
628
            except Exception as exc:
                raise AssertionError(
629
                    f"Transforming a PIL image with shape {image_repr} failed in the prototype transform with "
630
631
632
633
                    f"the error above. This means there is a consistency bug either in `_get_params` or in the "
                    f"`PIL.Image.Image` path in `_transform`."
                ) from exc

634
            assert_close(
635
636
                output_prototype_pil,
                output_legacy_pil,
637
                msg=lambda msg: f"PIL image consistency check failed with: \n\n{msg}",
638
                **closeness_kwargs,
639
            )
640
641


642
@pytest.mark.parametrize(
643
644
    ("config", "args_kwargs"),
    [
645
646
647
        pytest.param(
            config, args_kwargs, id=f"{config.legacy_cls.__name__}-{idx:0{len(str(len(config.args_kwargs)))}d}"
        )
648
        for config in CONSISTENCY_CONFIGS
649
        for idx, args_kwargs in enumerate(config.args_kwargs)
650
    ],
651
)
652
@pytest.mark.filterwarnings("ignore")
653
def test_call_consistency(config, args_kwargs):
654
655
656
    args, kwargs = args_kwargs

    try:
657
        legacy_transform = config.legacy_cls(*args, **kwargs)
658
659
660
661
662
663
664
    except Exception as exc:
        raise pytest.UsageError(
            f"Initializing the legacy transform failed with the error above. "
            f"Please correct the `ArgsKwargs({args_kwargs})` in the `ConsistencyConfig`."
        ) from exc

    try:
665
        prototype_transform = config.prototype_cls(*args, **kwargs)
666
667
668
669
670
671
    except Exception as exc:
        raise AssertionError(
            "Initializing the prototype transform failed with the error above. "
            "This means there is a consistency bug in the constructor."
        ) from exc

672
673
674
675
676
    check_call_consistency(
        prototype_transform,
        legacy_transform,
        images=make_images(**config.make_images_kwargs),
        supports_pil=config.supports_pil,
677
        closeness_kwargs=config.closeness_kwargs,
678
679
680
    )


681
682
683
684
685
686
687
688
689
get_params_parametrization = pytest.mark.parametrize(
    ("config", "get_params_args_kwargs"),
    [
        pytest.param(
            next(config for config in CONSISTENCY_CONFIGS if config.prototype_cls is transform_cls),
            get_params_args_kwargs,
            id=transform_cls.__name__,
        )
        for transform_cls, get_params_args_kwargs in [
690
691
692
693
694
            (v2_transforms.RandomResizedCrop, ArgsKwargs(make_image(), scale=[0.3, 0.7], ratio=[0.5, 1.5])),
            (v2_transforms.RandomErasing, ArgsKwargs(make_image(), scale=(0.3, 0.7), ratio=(0.5, 1.5))),
            (v2_transforms.ColorJitter, ArgsKwargs(brightness=None, contrast=None, saturation=None, hue=None)),
            (v2_transforms.ElasticTransform, ArgsKwargs(alpha=[15.3, 27.2], sigma=[2.5, 3.9], size=[17, 31])),
            (v2_transforms.GaussianBlur, ArgsKwargs(0.3, 1.4)),
695
            (
696
                v2_transforms.RandomAffine,
697
698
                ArgsKwargs(degrees=[-20.0, 10.0], translate=None, scale_ranges=None, shears=None, img_size=[15, 29]),
            ),
699
700
701
702
            (v2_transforms.RandomCrop, ArgsKwargs(make_image(size=(61, 47)), output_size=(19, 25))),
            (v2_transforms.RandomPerspective, ArgsKwargs(23, 17, 0.5)),
            (v2_transforms.RandomRotation, ArgsKwargs(degrees=[-20.0, 10.0])),
            (v2_transforms.AutoAugment, ArgsKwargs(5)),
703
704
        ]
    ],
705
)
706
707


708
@get_params_parametrization
709
def test_get_params_alias(config, get_params_args_kwargs):
710
711
    assert config.prototype_cls.get_params is config.legacy_cls.get_params

712
713
714
715
716
    if not config.args_kwargs:
        return
    args, kwargs = config.args_kwargs[0]
    legacy_transform = config.legacy_cls(*args, **kwargs)
    prototype_transform = config.prototype_cls(*args, **kwargs)
717

718
719
720
    assert prototype_transform.get_params is legacy_transform.get_params


721
@get_params_parametrization
722
723
724
725
726
727
728
729
730
def test_get_params_jit(config, get_params_args_kwargs):
    get_params_args, get_params_kwargs = get_params_args_kwargs

    torch.jit.script(config.prototype_cls.get_params)(*get_params_args, **get_params_kwargs)

    if not config.args_kwargs:
        return
    args, kwargs = config.args_kwargs[0]
    transform = config.prototype_cls(*args, **kwargs)
731

732
    torch.jit.script(transform.get_params)(*get_params_args, **get_params_kwargs)
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
764
765
766
@pytest.mark.parametrize(
    ("config", "args_kwargs"),
    [
        pytest.param(
            config, args_kwargs, id=f"{config.legacy_cls.__name__}-{idx:0{len(str(len(config.args_kwargs)))}d}"
        )
        for config in CONSISTENCY_CONFIGS
        for idx, args_kwargs in enumerate(config.args_kwargs)
        if not isinstance(args_kwargs, NotScriptableArgsKwargs)
    ],
)
def test_jit_consistency(config, args_kwargs):
    args, kwargs = args_kwargs

    prototype_transform_eager = config.prototype_cls(*args, **kwargs)
    legacy_transform_eager = config.legacy_cls(*args, **kwargs)

    legacy_transform_scripted = torch.jit.script(legacy_transform_eager)
    prototype_transform_scripted = torch.jit.script(prototype_transform_eager)

    for image in make_images(**config.make_images_kwargs):
        image = image.as_subclass(torch.Tensor)

        torch.manual_seed(0)
        output_legacy_scripted = legacy_transform_scripted(image)

        torch.manual_seed(0)
        output_prototype_scripted = prototype_transform_scripted(image)

        assert_close(output_prototype_scripted, output_legacy_scripted, **config.closeness_kwargs)


767
768
769
770
771
772
773
774
775
776
class TestContainerTransforms:
    """
    Since we are testing containers here, we also need some transforms to wrap. Thus, testing a container transform for
    consistency automatically tests the wrapped transforms consistency.

    Instead of complicated mocking or creating custom transforms just for these tests, here we use deterministic ones
    that were already tested for consistency above.
    """

    def test_compose(self):
777
        prototype_transform = v2_transforms.Compose(
778
            [
779
780
                v2_transforms.Resize(256),
                v2_transforms.CenterCrop(224),
781
782
783
784
785
786
787
788
789
            ]
        )
        legacy_transform = legacy_transforms.Compose(
            [
                legacy_transforms.Resize(256),
                legacy_transforms.CenterCrop(224),
            ]
        )

790
791
        # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
        check_call_consistency(prototype_transform, legacy_transform, closeness_kwargs=dict(rtol=0, atol=1))
792
793

    @pytest.mark.parametrize("p", [0, 0.1, 0.5, 0.9, 1])
794
795
    @pytest.mark.parametrize("sequence_type", [list, nn.ModuleList])
    def test_random_apply(self, p, sequence_type):
796
        prototype_transform = v2_transforms.RandomApply(
797
798
            sequence_type(
                [
799
800
                    v2_transforms.Resize(256),
                    v2_transforms.CenterCrop(224),
801
802
                ]
            ),
803
804
805
            p=p,
        )
        legacy_transform = legacy_transforms.RandomApply(
806
807
808
809
810
811
            sequence_type(
                [
                    legacy_transforms.Resize(256),
                    legacy_transforms.CenterCrop(224),
                ]
            ),
812
813
814
            p=p,
        )

815
816
        # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
        check_call_consistency(prototype_transform, legacy_transform, closeness_kwargs=dict(rtol=0, atol=1))
817

818
819
820
821
822
        if sequence_type is nn.ModuleList:
            # quick and dirty test that it is jit-scriptable
            scripted = torch.jit.script(prototype_transform)
            scripted(torch.rand(1, 3, 300, 300))

823
    # We can't test other values for `p` since the random parameter generation is different
824
825
    @pytest.mark.parametrize("probabilities", [(0, 1), (1, 0)])
    def test_random_choice(self, probabilities):
826
        prototype_transform = v2_transforms.RandomChoice(
827
            [
828
                v2_transforms.Resize(256),
829
830
                legacy_transforms.CenterCrop(224),
            ],
831
            p=probabilities,
832
833
834
835
836
837
        )
        legacy_transform = legacy_transforms.RandomChoice(
            [
                legacy_transforms.Resize(256),
                legacy_transforms.CenterCrop(224),
            ],
838
            p=probabilities,
839
840
        )

841
842
        # atol=1 due to Resize v2 is using native uint8 interpolate path for bilinear and nearest modes
        check_call_consistency(prototype_transform, legacy_transform, closeness_kwargs=dict(rtol=0, atol=1))
843
844


845
846
class TestToTensorTransforms:
    def test_pil_to_tensor(self):
847
        prototype_transform = v2_transforms.PILToTensor()
848
849
        legacy_transform = legacy_transforms.PILToTensor()

850
851
852
853
854
855
        for image in make_images(extra_dims=[()]):
            image_pil = to_image_pil(image)

            assert_equal(prototype_transform(image_pil), legacy_transform(image_pil))

    def test_to_tensor(self):
856
        with pytest.warns(UserWarning, match=re.escape("The transform `ToTensor()` is deprecated")):
857
            prototype_transform = v2_transforms.ToTensor()
858
859
        legacy_transform = legacy_transforms.ToTensor()

860
861
862
863
864
865
        for image in make_images(extra_dims=[()]):
            image_pil = to_image_pil(image)
            image_numpy = np.array(image_pil)

            assert_equal(prototype_transform(image_pil), legacy_transform(image_pil))
            assert_equal(prototype_transform(image_numpy), legacy_transform(image_numpy))
866
867
868
869
870
871
872
873


class TestAATransforms:
    @pytest.mark.parametrize(
        "inpt",
        [
            torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8),
            PIL.Image.new("RGB", (256, 256), 123),
874
            datapoints.Image(torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8)),
875
876
877
878
        ],
    )
    @pytest.mark.parametrize(
        "interpolation",
879
        [
880
881
            v2_transforms.InterpolationMode.NEAREST,
            v2_transforms.InterpolationMode.BILINEAR,
882
883
            PIL.Image.NEAREST,
        ],
884
885
886
    )
    def test_randaug(self, inpt, interpolation, mocker):
        t_ref = legacy_transforms.RandAugment(interpolation=interpolation, num_ops=1)
887
        t = v2_transforms.RandAugment(interpolation=interpolation, num_ops=1)
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908

        le = len(t._AUGMENTATION_SPACE)
        keys = list(t._AUGMENTATION_SPACE.keys())
        randint_values = []
        for i in range(le):
            # Stable API, op_index random call
            randint_values.append(i)
            # Stable API, if signed there is another random call
            if t._AUGMENTATION_SPACE[keys[i]][1]:
                randint_values.append(0)
            # New API, _get_random_item
            randint_values.append(i)
        randint_values = iter(randint_values)

        mocker.patch("torch.randint", side_effect=lambda *arg, **kwargs: torch.tensor(next(randint_values)))
        mocker.patch("torch.rand", return_value=1.0)

        for i in range(le):
            expected_output = t_ref(inpt)
            output = t(inpt)

909
            assert_close(expected_output, output, atol=1, rtol=0.1)
910
911
912
913
914
915

    @pytest.mark.parametrize(
        "inpt",
        [
            torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8),
            PIL.Image.new("RGB", (256, 256), 123),
916
            datapoints.Image(torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8)),
917
918
919
920
        ],
    )
    @pytest.mark.parametrize(
        "interpolation",
921
        [
922
923
            v2_transforms.InterpolationMode.NEAREST,
            v2_transforms.InterpolationMode.BILINEAR,
924
925
            PIL.Image.NEAREST,
        ],
926
927
928
    )
    def test_trivial_aug(self, inpt, interpolation, mocker):
        t_ref = legacy_transforms.TrivialAugmentWide(interpolation=interpolation)
929
        t = v2_transforms.TrivialAugmentWide(interpolation=interpolation)
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960

        le = len(t._AUGMENTATION_SPACE)
        keys = list(t._AUGMENTATION_SPACE.keys())
        randint_values = []
        for i in range(le):
            # Stable API, op_index random call
            randint_values.append(i)
            key = keys[i]
            # Stable API, random magnitude
            aug_op = t._AUGMENTATION_SPACE[key]
            magnitudes = aug_op[0](2, 0, 0)
            if magnitudes is not None:
                randint_values.append(5)
            # Stable API, if signed there is another random call
            if aug_op[1]:
                randint_values.append(0)
            # New API, _get_random_item
            randint_values.append(i)
            # New API, random magnitude
            if magnitudes is not None:
                randint_values.append(5)

        randint_values = iter(randint_values)

        mocker.patch("torch.randint", side_effect=lambda *arg, **kwargs: torch.tensor(next(randint_values)))
        mocker.patch("torch.rand", return_value=1.0)

        for _ in range(le):
            expected_output = t_ref(inpt)
            output = t(inpt)

961
            assert_close(expected_output, output, atol=1, rtol=0.1)
962
963
964
965
966
967

    @pytest.mark.parametrize(
        "inpt",
        [
            torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8),
            PIL.Image.new("RGB", (256, 256), 123),
968
            datapoints.Image(torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8)),
969
970
971
972
        ],
    )
    @pytest.mark.parametrize(
        "interpolation",
973
        [
974
975
            v2_transforms.InterpolationMode.NEAREST,
            v2_transforms.InterpolationMode.BILINEAR,
976
977
            PIL.Image.NEAREST,
        ],
978
979
980
981
    )
    def test_augmix(self, inpt, interpolation, mocker):
        t_ref = legacy_transforms.AugMix(interpolation=interpolation, mixture_width=1, chain_depth=1)
        t_ref._sample_dirichlet = lambda t: t.softmax(dim=-1)
982
        t = v2_transforms.AugMix(interpolation=interpolation, mixture_width=1, chain_depth=1)
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
        t._sample_dirichlet = lambda t: t.softmax(dim=-1)

        le = len(t._AUGMENTATION_SPACE)
        keys = list(t._AUGMENTATION_SPACE.keys())
        randint_values = []
        for i in range(le):
            # Stable API, op_index random call
            randint_values.append(i)
            key = keys[i]
            # Stable API, random magnitude
            aug_op = t._AUGMENTATION_SPACE[key]
            magnitudes = aug_op[0](2, 0, 0)
            if magnitudes is not None:
                randint_values.append(5)
            # Stable API, if signed there is another random call
            if aug_op[1]:
                randint_values.append(0)
            # New API, _get_random_item
            randint_values.append(i)
            # New API, random magnitude
            if magnitudes is not None:
                randint_values.append(5)

        randint_values = iter(randint_values)

        mocker.patch("torch.randint", side_effect=lambda *arg, **kwargs: torch.tensor(next(randint_values)))
        mocker.patch("torch.rand", return_value=1.0)

        expected_output = t_ref(inpt)
        output = t(inpt)

        assert_equal(expected_output, output)

    @pytest.mark.parametrize(
        "inpt",
        [
            torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8),
            PIL.Image.new("RGB", (256, 256), 123),
1021
            datapoints.Image(torch.randint(0, 256, size=(1, 3, 256, 256), dtype=torch.uint8)),
1022
1023
1024
1025
        ],
    )
    @pytest.mark.parametrize(
        "interpolation",
1026
        [
1027
1028
            v2_transforms.InterpolationMode.NEAREST,
            v2_transforms.InterpolationMode.BILINEAR,
1029
1030
            PIL.Image.NEAREST,
        ],
1031
1032
1033
1034
    )
    def test_aa(self, inpt, interpolation):
        aa_policy = legacy_transforms.AutoAugmentPolicy("imagenet")
        t_ref = legacy_transforms.AutoAugment(aa_policy, interpolation=interpolation)
1035
        t = v2_transforms.AutoAugment(aa_policy, interpolation=interpolation)
1036
1037
1038
1039
1040
1041
1042
1043

        torch.manual_seed(12)
        expected_output = t_ref(inpt)

        torch.manual_seed(12)
        output = t(inpt)

        assert_equal(expected_output, output)
1044
1045


1046
def import_transforms_from_references(reference):
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
    HERE = Path(__file__).parent
    PROJECT_ROOT = HERE.parent

    loader = importlib.machinery.SourceFileLoader(
        "transforms", str(PROJECT_ROOT / "references" / reference / "transforms.py")
    )
    spec = importlib.util.spec_from_loader("transforms", loader)
    module = importlib.util.module_from_spec(spec)
    loader.exec_module(module)
    return module
1057
1058
1059


det_transforms = import_transforms_from_references("detection")
1060
1061
1062
1063
1064
1065
1066


class TestRefDetTransforms:
    def make_datapoints(self, with_mask=True):
        size = (600, 800)
        num_objects = 22

1067
1068
1069
        def make_label(extra_dims, categories):
            return torch.randint(categories, extra_dims, dtype=torch.int64)

1070
        pil_image = to_image_pil(make_image(size=size, color_space="RGB"))
1071
        target = {
1072
            "boxes": make_bounding_box(spatial_size=size, format="XYXY", extra_dims=(num_objects,), dtype=torch.float),
1073
1074
1075
1076
1077
1078
1079
            "labels": make_label(extra_dims=(num_objects,), categories=80),
        }
        if with_mask:
            target["masks"] = make_detection_mask(size=size, num_objects=num_objects, dtype=torch.long)

        yield (pil_image, target)

1080
        tensor_image = torch.Tensor(make_image(size=size, color_space="RGB"))
1081
        target = {
1082
            "boxes": make_bounding_box(spatial_size=size, format="XYXY", extra_dims=(num_objects,), dtype=torch.float),
1083
1084
1085
1086
1087
1088
1089
            "labels": make_label(extra_dims=(num_objects,), categories=80),
        }
        if with_mask:
            target["masks"] = make_detection_mask(size=size, num_objects=num_objects, dtype=torch.long)

        yield (tensor_image, target)

1090
        datapoint_image = make_image(size=size, color_space="RGB")
1091
        target = {
1092
            "boxes": make_bounding_box(spatial_size=size, format="XYXY", extra_dims=(num_objects,), dtype=torch.float),
1093
1094
1095
1096
1097
            "labels": make_label(extra_dims=(num_objects,), categories=80),
        }
        if with_mask:
            target["masks"] = make_detection_mask(size=size, num_objects=num_objects, dtype=torch.long)

1098
        yield (datapoint_image, target)
1099
1100
1101
1102

    @pytest.mark.parametrize(
        "t_ref, t, data_kwargs",
        [
1103
            (det_transforms.RandomHorizontalFlip(p=1.0), v2_transforms.RandomHorizontalFlip(p=1.0), {}),
1104
1105
1106
1107
1108
            (
                det_transforms.RandomIoUCrop(),
                v2_transforms.Compose(
                    [
                        v2_transforms.RandomIoUCrop(),
1109
                        v2_transforms.SanitizeBoundingBox(labels_getter=lambda sample: sample[1]["labels"]),
1110
1111
1112
1113
                    ]
                ),
                {"with_mask": False},
            ),
1114
1115
            (det_transforms.RandomZoomOut(), v2_transforms.RandomZoomOut(), {"with_mask": False}),
            (det_transforms.ScaleJitter((1024, 1024)), v2_transforms.ScaleJitter((1024, 1024)), {}),
1116
1117
1118
1119
            (
                det_transforms.RandomShortestSize(
                    min_size=(480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800), max_size=1333
                ),
1120
                v2_transforms.RandomShortestSize(
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
                    min_size=(480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800), max_size=1333
                ),
                {},
            ),
        ],
    )
    def test_transform(self, t_ref, t, data_kwargs):
        for dp in self.make_datapoints(**data_kwargs):

            # We should use prototype transform first as reference transform performs inplace target update
            torch.manual_seed(12)
            output = t(dp)

            torch.manual_seed(12)
            expected_output = t_ref(*dp)

            assert_equal(expected_output, output)
1138
1139
1140
1141
1142
1143
1144
1145
1146


seg_transforms = import_transforms_from_references("segmentation")


# We need this transform for two reasons:
# 1. transforms.RandomCrop uses a different scheme to pad images and masks of insufficient size than its name
#    counterpart in the detection references. Thus, we cannot use it with `pad_if_needed=True`
# 2. transforms.Pad only supports a fixed padding, but the segmentation datasets don't have a fixed image size.
1147
class PadIfSmaller(v2_transforms.Transform):
1148
1149
1150
    def __init__(self, size, fill=0):
        super().__init__()
        self.size = size
1151
        self.fill = v2_transforms._geometry._setup_fill_arg(fill)
1152
1153

    def _get_params(self, sample):
1154
        height, width = query_spatial_size(sample)
1155
1156
1157
1158
1159
1160
1161
1162
1163
        padding = [0, 0, max(self.size - width, 0), max(self.size - height, 0)]
        needs_padding = any(padding)
        return dict(padding=padding, needs_padding=needs_padding)

    def _transform(self, inpt, params):
        if not params["needs_padding"]:
            return inpt

        fill = self.fill[type(inpt)]
1164
        return prototype_F.pad(inpt, padding=params["padding"], fill=fill)
1165
1166
1167
1168


class TestRefSegTransforms:
    def make_datapoints(self, supports_pil=True, image_dtype=torch.uint8):
1169
        size = (256, 460)
1170
1171
1172
1173
1174
1175
1176
1177
        num_categories = 21

        conv_fns = []
        if supports_pil:
            conv_fns.append(to_image_pil)
        conv_fns.extend([torch.Tensor, lambda x: x])

        for conv_fn in conv_fns:
1178
            datapoint_image = make_image(size=size, color_space="RGB", dtype=image_dtype)
1179
            datapoint_mask = make_segmentation_mask(size=size, num_categories=num_categories, dtype=torch.uint8)
1180

1181
            dp = (conv_fn(datapoint_image), datapoint_mask)
1182
            dp_ref = (
1183
1184
                to_image_pil(datapoint_image) if supports_pil else datapoint_image.as_subclass(torch.Tensor),
                to_image_pil(datapoint_mask),
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
            )

            yield dp, dp_ref

    def set_seed(self, seed=12):
        torch.manual_seed(seed)
        random.seed(seed)

    def check(self, t, t_ref, data_kwargs=None):
        for dp, dp_ref in self.make_datapoints(**data_kwargs or dict()):

            self.set_seed()
1197
            actual = actual_image, actual_mask = t(dp)
1198
1199

            self.set_seed()
1200
1201
1202
1203
1204
            expected_image, expected_mask = t_ref(*dp_ref)
            if isinstance(actual_image, torch.Tensor) and not isinstance(expected_image, torch.Tensor):
                expected_image = legacy_F.pil_to_tensor(expected_image)
            expected_mask = legacy_F.pil_to_tensor(expected_mask).squeeze(0)
            expected = (expected_image, expected_mask)
1205

1206
            assert_equal(actual, expected)
1207
1208
1209
1210
1211
1212

    @pytest.mark.parametrize(
        ("t_ref", "t", "data_kwargs"),
        [
            (
                seg_transforms.RandomHorizontalFlip(flip_prob=1.0),
1213
                v2_transforms.RandomHorizontalFlip(p=1.0),
1214
1215
1216
1217
                dict(),
            ),
            (
                seg_transforms.RandomHorizontalFlip(flip_prob=0.0),
1218
                v2_transforms.RandomHorizontalFlip(p=0.0),
1219
1220
1221
1222
                dict(),
            ),
            (
                seg_transforms.RandomCrop(size=480),
1223
                v2_transforms.Compose(
1224
                    [
1225
                        PadIfSmaller(size=480, fill=defaultdict(lambda: 0, {datapoints.Mask: 255})),
1226
                        v2_transforms.RandomCrop(size=480),
1227
1228
1229
1230
1231
1232
                    ]
                ),
                dict(),
            ),
            (
                seg_transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
1233
                v2_transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
1234
1235
1236
1237
1238
1239
1240
1241
                dict(supports_pil=False, image_dtype=torch.float),
            ),
        ],
    )
    def test_common(self, t_ref, t, data_kwargs):
        self.check(t, t_ref, data_kwargs)

    def check_resize(self, mocker, t_ref, t):
1242
        mock = mocker.patch("torchvision.transforms.v2._geometry.F.resize")
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
        mock_ref = mocker.patch("torchvision.transforms.functional.resize")

        for dp, dp_ref in self.make_datapoints():
            mock.reset_mock()
            mock_ref.reset_mock()

            self.set_seed()
            t(dp)
            assert mock.call_count == 2
            assert all(
                actual is expected
                for actual, expected in zip([call_args[0][0] for call_args in mock.call_args_list], dp)
            )

            self.set_seed()
            t_ref(*dp_ref)
            assert mock_ref.call_count == 2
            assert all(
                actual is expected
                for actual, expected in zip([call_args[0][0] for call_args in mock_ref.call_args_list], dp_ref)
            )

            for args_kwargs, args_kwargs_ref in zip(mock.call_args_list, mock_ref.call_args_list):
                assert args_kwargs[0][1] == [args_kwargs_ref[0][1]]

    def test_random_resize_train(self, mocker):
        base_size = 520
        min_size = base_size // 2
        max_size = base_size * 2

        randint = torch.randint

        def patched_randint(a, b, *other_args, **kwargs):
            if kwargs or len(other_args) > 1 or other_args[0] != ():
                return randint(a, b, *other_args, **kwargs)

            return random.randint(a, b)

        # We are patching torch.randint -> random.randint here, because we can't patch the modules that are not imported
        # normally
1283
        t = v2_transforms.RandomResize(min_size=min_size, max_size=max_size, antialias=True)
1284
        mocker.patch(
1285
            "torchvision.transforms.v2._geometry.torch.randint",
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
            new=patched_randint,
        )

        t_ref = seg_transforms.RandomResize(min_size=min_size, max_size=max_size)

        self.check_resize(mocker, t_ref, t)

    def test_random_resize_eval(self, mocker):
        torch.manual_seed(0)
        base_size = 520

1297
        t = v2_transforms.Resize(size=base_size, antialias=True)
1298
1299
1300
1301

        t_ref = seg_transforms.RandomResize(min_size=base_size, max_size=base_size)

        self.check_resize(mocker, t_ref, t)
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314


@pytest.mark.parametrize(
    ("legacy_dispatcher", "name_only_params"),
    [
        (legacy_F.get_dimensions, {}),
        (legacy_F.get_image_size, {}),
        (legacy_F.get_image_num_channels, {}),
        (legacy_F.to_tensor, {}),
        (legacy_F.pil_to_tensor, {}),
        (legacy_F.convert_image_dtype, {}),
        (legacy_F.to_pil_image, {}),
        (legacy_F.normalize, {}),
1315
        (legacy_F.resize, {"interpolation"}),
1316
1317
1318
        (legacy_F.pad, {"padding", "fill"}),
        (legacy_F.crop, {}),
        (legacy_F.center_crop, {}),
1319
        (legacy_F.resized_crop, {"interpolation"}),
1320
        (legacy_F.hflip, {}),
1321
        (legacy_F.perspective, {"startpoints", "endpoints", "fill", "interpolation"}),
1322
1323
1324
1325
1326
1327
1328
1329
        (legacy_F.vflip, {}),
        (legacy_F.five_crop, {}),
        (legacy_F.ten_crop, {}),
        (legacy_F.adjust_brightness, {}),
        (legacy_F.adjust_contrast, {}),
        (legacy_F.adjust_saturation, {}),
        (legacy_F.adjust_hue, {}),
        (legacy_F.adjust_gamma, {}),
1330
1331
        (legacy_F.rotate, {"center", "fill", "interpolation"}),
        (legacy_F.affine, {"angle", "translate", "center", "fill", "interpolation"}),
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
        (legacy_F.to_grayscale, {}),
        (legacy_F.rgb_to_grayscale, {}),
        (legacy_F.to_tensor, {}),
        (legacy_F.erase, {}),
        (legacy_F.gaussian_blur, {}),
        (legacy_F.invert, {}),
        (legacy_F.posterize, {}),
        (legacy_F.solarize, {}),
        (legacy_F.adjust_sharpness, {}),
        (legacy_F.autocontrast, {}),
        (legacy_F.equalize, {}),
1343
        (legacy_F.elastic_transform, {"fill", "interpolation"}),
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
    ],
)
def test_dispatcher_signature_consistency(legacy_dispatcher, name_only_params):
    legacy_signature = inspect.signature(legacy_dispatcher)
    legacy_params = list(legacy_signature.parameters.values())[1:]

    try:
        prototype_dispatcher = getattr(prototype_F, legacy_dispatcher.__name__)
    except AttributeError:
        raise AssertionError(
            f"Legacy dispatcher `F.{legacy_dispatcher.__name__}` has no prototype equivalent"
        ) from None

    prototype_signature = inspect.signature(prototype_dispatcher)
    prototype_params = list(prototype_signature.parameters.values())[1:]

    # Some dispatchers got extra parameters. This makes sure they have a default argument and thus are BC. We don't
    # need to check if parameters were added in the middle rather than at the end, since that will be caught by the
    # regular check below.
    prototype_params, new_prototype_params = (
        prototype_params[: len(legacy_params)],
        prototype_params[len(legacy_params) :],
    )
    for param in new_prototype_params:
        assert param.default is not param.empty

    # Some annotations were changed mostly to supersets of what was there before. Plus, some legacy dispatchers had no
    # annotations. In these cases we simply drop the annotation and default argument from the comparison
    for prototype_param, legacy_param in zip(prototype_params, legacy_params):
        if legacy_param.name in name_only_params:
            prototype_param._annotation = prototype_param._default = inspect.Parameter.empty
            legacy_param._annotation = legacy_param._default = inspect.Parameter.empty
        elif legacy_param.annotation is inspect.Parameter.empty:
            prototype_param._annotation = inspect.Parameter.empty

    assert prototype_params == legacy_params