_geometry.py 84.5 KB
Newer Older
1
import math
2
import numbers
3
import warnings
4
from typing import Any, List, Optional, Sequence, Tuple, Union
5
6
7

import PIL.Image
import torch
8
from torch.nn.functional import grid_sample, interpolate, pad as torch_pad
9

10
from torchvision import datapoints
11
12
from torchvision.transforms import _functional_pil as _FP
from torchvision.transforms._functional_tensor import _pad_symmetric
13
from torchvision.transforms.functional import (
14
    _check_antialias,
15
    _compute_resized_output_size as __compute_resized_output_size,
16
    _get_perspective_coeffs,
17
    _interpolation_modes_from_int,
18
    InterpolationMode,
19
    pil_modes_mapping,
20
21
    pil_to_tensor,
    to_pil_image,
22
)
23

24
25
from torchvision.utils import _log_api_usage_once

Philip Meier's avatar
Philip Meier committed
26
from ._meta import clamp_bounding_boxes, convert_format_bounding_boxes, get_size_image_pil
27

28
from ._utils import _FillTypeJIT, _get_kernel, _register_five_ten_crop_kernel_internal, _register_kernel_internal
29

30

31
32
33
34
35
36
37
38
39
40
41
def _check_interpolation(interpolation: Union[InterpolationMode, int]) -> InterpolationMode:
    if isinstance(interpolation, int):
        interpolation = _interpolation_modes_from_int(interpolation)
    elif not isinstance(interpolation, InterpolationMode):
        raise ValueError(
            f"Argument interpolation should be an `InterpolationMode` or a corresponding Pillow integer constant, "
            f"but got {interpolation}."
        )
    return interpolation


42
def horizontal_flip(inpt: torch.Tensor) -> torch.Tensor:
43
    if torch.jit.is_scripting():
44
        return horizontal_flip_image_tensor(inpt)
45
46
47
48
49

    _log_api_usage_once(horizontal_flip)

    kernel = _get_kernel(horizontal_flip, type(inpt))
    return kernel(inpt)
50
51


52
@_register_kernel_internal(horizontal_flip, torch.Tensor)
53
@_register_kernel_internal(horizontal_flip, datapoints.Image)
54
55
56
57
def horizontal_flip_image_tensor(image: torch.Tensor) -> torch.Tensor:
    return image.flip(-1)


58
@_register_kernel_internal(horizontal_flip, PIL.Image.Image)
59
60
def horizontal_flip_image_pil(image: PIL.Image.Image) -> PIL.Image.Image:
    return _FP.hflip(image)
61
62


63
@_register_kernel_internal(horizontal_flip, datapoints.Mask)
64
65
def horizontal_flip_mask(mask: torch.Tensor) -> torch.Tensor:
    return horizontal_flip_image_tensor(mask)
66
67


68
def horizontal_flip_bounding_boxes(
Philip Meier's avatar
Philip Meier committed
69
    bounding_boxes: torch.Tensor, format: datapoints.BoundingBoxFormat, canvas_size: Tuple[int, int]
70
) -> torch.Tensor:
71
    shape = bounding_boxes.shape
72

73
    bounding_boxes = bounding_boxes.clone().reshape(-1, 4)
74

75
    if format == datapoints.BoundingBoxFormat.XYXY:
Philip Meier's avatar
Philip Meier committed
76
        bounding_boxes[:, [2, 0]] = bounding_boxes[:, [0, 2]].sub_(canvas_size[1]).neg_()
77
    elif format == datapoints.BoundingBoxFormat.XYWH:
Philip Meier's avatar
Philip Meier committed
78
        bounding_boxes[:, 0].add_(bounding_boxes[:, 2]).sub_(canvas_size[1]).neg_()
79
    else:  # format == datapoints.BoundingBoxFormat.CXCYWH:
Philip Meier's avatar
Philip Meier committed
80
        bounding_boxes[:, 0].sub_(canvas_size[1]).neg_()
81

82
    return bounding_boxes.reshape(shape)
83
84


85
86
87
88
89
@_register_kernel_internal(horizontal_flip, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _horizontal_flip_bounding_boxes_dispatch(inpt: datapoints.BoundingBoxes) -> datapoints.BoundingBoxes:
    output = horizontal_flip_bounding_boxes(
        inpt.as_subclass(torch.Tensor), format=inpt.format, canvas_size=inpt.canvas_size
    )
90
    return datapoints.wrap(output, like=inpt)
91
92
93


@_register_kernel_internal(horizontal_flip, datapoints.Video)
94
95
96
97
def horizontal_flip_video(video: torch.Tensor) -> torch.Tensor:
    return horizontal_flip_image_tensor(video)


98
def vertical_flip(inpt: torch.Tensor) -> torch.Tensor:
99
    if torch.jit.is_scripting():
100
        return vertical_flip_image_tensor(inpt)
101
102
103
104
105

    _log_api_usage_once(vertical_flip)

    kernel = _get_kernel(vertical_flip, type(inpt))
    return kernel(inpt)
106
107


108
@_register_kernel_internal(vertical_flip, torch.Tensor)
109
@_register_kernel_internal(vertical_flip, datapoints.Image)
110
111
112
113
def vertical_flip_image_tensor(image: torch.Tensor) -> torch.Tensor:
    return image.flip(-2)


114
@_register_kernel_internal(vertical_flip, PIL.Image.Image)
Philip Meier's avatar
Philip Meier committed
115
116
def vertical_flip_image_pil(image: PIL.Image) -> PIL.Image:
    return _FP.vflip(image)
117
118


119
@_register_kernel_internal(vertical_flip, datapoints.Mask)
120
121
def vertical_flip_mask(mask: torch.Tensor) -> torch.Tensor:
    return vertical_flip_image_tensor(mask)
122
123


124
def vertical_flip_bounding_boxes(
Philip Meier's avatar
Philip Meier committed
125
    bounding_boxes: torch.Tensor, format: datapoints.BoundingBoxFormat, canvas_size: Tuple[int, int]
126
) -> torch.Tensor:
127
    shape = bounding_boxes.shape
128

129
    bounding_boxes = bounding_boxes.clone().reshape(-1, 4)
130

131
    if format == datapoints.BoundingBoxFormat.XYXY:
Philip Meier's avatar
Philip Meier committed
132
        bounding_boxes[:, [1, 3]] = bounding_boxes[:, [3, 1]].sub_(canvas_size[0]).neg_()
133
    elif format == datapoints.BoundingBoxFormat.XYWH:
Philip Meier's avatar
Philip Meier committed
134
        bounding_boxes[:, 1].add_(bounding_boxes[:, 3]).sub_(canvas_size[0]).neg_()
135
    else:  # format == datapoints.BoundingBoxFormat.CXCYWH:
Philip Meier's avatar
Philip Meier committed
136
        bounding_boxes[:, 1].sub_(canvas_size[0]).neg_()
137

138
    return bounding_boxes.reshape(shape)
139
140


141
142
143
144
145
@_register_kernel_internal(vertical_flip, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _vertical_flip_bounding_boxes_dispatch(inpt: datapoints.BoundingBoxes) -> datapoints.BoundingBoxes:
    output = vertical_flip_bounding_boxes(
        inpt.as_subclass(torch.Tensor), format=inpt.format, canvas_size=inpt.canvas_size
    )
146
    return datapoints.wrap(output, like=inpt)
147

148

149
150
151
@_register_kernel_internal(vertical_flip, datapoints.Video)
def vertical_flip_video(video: torch.Tensor) -> torch.Tensor:
    return vertical_flip_image_tensor(video)
152
153


154
155
156
157
158
159
# We changed the names to align them with the transforms, i.e. `RandomHorizontalFlip`. Still, `hflip` and `vflip` are
# prevalent and well understood. Thus, we just alias them without deprecating the old names.
hflip = horizontal_flip
vflip = vertical_flip


160
def _compute_resized_output_size(
Philip Meier's avatar
Philip Meier committed
161
    canvas_size: Tuple[int, int], size: List[int], max_size: Optional[int] = None
162
163
164
) -> List[int]:
    if isinstance(size, int):
        size = [size]
165
166
167
168
169
    elif max_size is not None and len(size) != 1:
        raise ValueError(
            "max_size should only be passed if size specifies the length of the smaller edge, "
            "i.e. size should be an int or a sequence of length 1 in torchscript mode."
        )
Philip Meier's avatar
Philip Meier committed
170
    return __compute_resized_output_size(canvas_size, size=size, max_size=max_size)
171
172


173
def resize(
174
    inpt: torch.Tensor,
175
176
177
178
    size: List[int],
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
    max_size: Optional[int] = None,
    antialias: Optional[Union[str, bool]] = "warn",
179
) -> torch.Tensor:
180
181
182
183
184
185
186
    if torch.jit.is_scripting():
        return resize_image_tensor(inpt, size=size, interpolation=interpolation, max_size=max_size, antialias=antialias)

    _log_api_usage_once(resize)

    kernel = _get_kernel(resize, type(inpt))
    return kernel(inpt, size=size, interpolation=interpolation, max_size=max_size, antialias=antialias)
187
188


189
@_register_kernel_internal(resize, torch.Tensor)
190
@_register_kernel_internal(resize, datapoints.Image)
191
192
193
def resize_image_tensor(
    image: torch.Tensor,
    size: List[int],
194
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
195
    max_size: Optional[int] = None,
196
    antialias: Optional[Union[str, bool]] = "warn",
197
) -> torch.Tensor:
198
    interpolation = _check_interpolation(interpolation)
199
200
    antialias = _check_antialias(img=image, antialias=antialias, interpolation=interpolation)
    assert not isinstance(antialias, str)
201
    antialias = False if antialias is None else antialias
202
203
204
    align_corners: Optional[bool] = None
    if interpolation == InterpolationMode.BILINEAR or interpolation == InterpolationMode.BICUBIC:
        align_corners = False
205
206
207
208
    else:
        # The default of antialias should be True from 0.17, so we don't warn or
        # error if other interpolation modes are used. This is documented.
        antialias = False
209

210
    shape = image.shape
211
    numel = image.numel()
212
    num_channels, old_height, old_width = shape[-3:]
vfdev's avatar
vfdev committed
213
    new_height, new_width = _compute_resized_output_size((old_height, old_width), size=size, max_size=max_size)
214

215
216
    if (new_height, new_width) == (old_height, old_width):
        return image
217
    elif numel > 0:
218
        image = image.reshape(-1, num_channels, old_height, old_width)
219

220
        dtype = image.dtype
221
222
223
224
        acceptable_dtypes = [torch.float32, torch.float64]
        if interpolation == InterpolationMode.NEAREST or interpolation == InterpolationMode.NEAREST_EXACT:
            # uint8 dtype can be included for cpu and cuda input if nearest mode
            acceptable_dtypes.append(torch.uint8)
225
226
227
228
229
230
231
        elif image.device.type == "cpu":
            # uint8 dtype support for bilinear and bicubic is limited to cpu and
            # according to our benchmarks, non-AVX CPUs should still prefer u8->f32->interpolate->u8 path for bilinear
            if (interpolation == InterpolationMode.BILINEAR and "AVX2" in torch.backends.cpu.get_cpu_capability()) or (
                interpolation == InterpolationMode.BICUBIC
            ):
                acceptable_dtypes.append(torch.uint8)
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247

        strides = image.stride()
        if image.is_contiguous(memory_format=torch.channels_last) and image.shape[0] == 1 and numel != strides[0]:
            # There is a weird behaviour in torch core where the output tensor of `interpolate()` can be allocated as
            # contiguous even though the input is un-ambiguously channels_last (https://github.com/pytorch/pytorch/issues/68430).
            # In particular this happens for the typical torchvision use-case of single CHW images where we fake the batch dim
            # to become 1CHW. Below, we restride those tensors to trick torch core into properly allocating the output as
            # channels_last, thus preserving the memory format of the input. This is not just for format consistency:
            # for uint8 bilinear images, this also avoids an extra copy (re-packing) of the output and saves time.
            # TODO: when https://github.com/pytorch/pytorch/issues/68430 is fixed (possibly by https://github.com/pytorch/pytorch/pull/100373),
            # we should be able to remove this hack.
            new_strides = list(strides)
            new_strides[0] = numel
            image = image.as_strided((1, num_channels, old_height, old_width), new_strides)

        need_cast = dtype not in acceptable_dtypes
248
249
250
251
        if need_cast:
            image = image.to(dtype=torch.float32)

        image = interpolate(
252
253
            image,
            size=[new_height, new_width],
254
255
            mode=interpolation.value,
            align_corners=align_corners,
256
257
            antialias=antialias,
        )
258

259
260
        if need_cast:
            if interpolation == InterpolationMode.BICUBIC and dtype == torch.uint8:
261
                # This path is hit on non-AVX archs, or on GPU.
262
                image = image.clamp_(min=0, max=255)
263
264
265
            if dtype in (torch.uint8, torch.int8, torch.int16, torch.int32, torch.int64):
                image = image.round_()
            image = image.to(dtype=dtype)
266

267
    return image.reshape(shape[:-3] + (num_channels, new_height, new_width))
268
269
270


def resize_image_pil(
271
    image: PIL.Image.Image,
272
    size: Union[Sequence[int], int],
273
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
274
275
    max_size: Optional[int] = None,
) -> PIL.Image.Image:
276
277
278
279
280
281
282
    old_height, old_width = image.height, image.width
    new_height, new_width = _compute_resized_output_size(
        (old_height, old_width),
        size=size,  # type: ignore[arg-type]
        max_size=max_size,
    )

283
    interpolation = _check_interpolation(interpolation)
284
285
286
287
288

    if (new_height, new_width) == (old_height, old_width):
        return image

    return image.resize((new_width, new_height), resample=pil_modes_mapping[interpolation])
289
290


291
292
293
294
295
296
297
298
299
300
301
302
303
@_register_kernel_internal(resize, PIL.Image.Image)
def _resize_image_pil_dispatch(
    image: PIL.Image.Image,
    size: Union[Sequence[int], int],
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
    max_size: Optional[int] = None,
    antialias: Optional[Union[str, bool]] = "warn",
) -> PIL.Image.Image:
    if antialias is False:
        warnings.warn("Anti-alias option is always applied for PIL Image input. Argument antialias is ignored.")
    return resize_image_pil(image, size=size, interpolation=interpolation, max_size=max_size)


304
305
306
def resize_mask(mask: torch.Tensor, size: List[int], max_size: Optional[int] = None) -> torch.Tensor:
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
307
308
309
310
        needs_squeeze = True
    else:
        needs_squeeze = False

311
    output = resize_image_tensor(mask, size=size, interpolation=InterpolationMode.NEAREST, max_size=max_size)
312
313
314
315
316

    if needs_squeeze:
        output = output.squeeze(0)

    return output
317
318


319
320
321
322
323
@_register_kernel_internal(resize, datapoints.Mask, datapoint_wrapper=False)
def _resize_mask_dispatch(
    inpt: datapoints.Mask, size: List[int], max_size: Optional[int] = None, **kwargs: Any
) -> datapoints.Mask:
    output = resize_mask(inpt.as_subclass(torch.Tensor), size, max_size=max_size)
324
    return datapoints.wrap(output, like=inpt)
325
326


327
def resize_bounding_boxes(
Philip Meier's avatar
Philip Meier committed
328
    bounding_boxes: torch.Tensor, canvas_size: Tuple[int, int], size: List[int], max_size: Optional[int] = None
329
) -> Tuple[torch.Tensor, Tuple[int, int]]:
Philip Meier's avatar
Philip Meier committed
330
331
    old_height, old_width = canvas_size
    new_height, new_width = _compute_resized_output_size(canvas_size, size=size, max_size=max_size)
332
333

    if (new_height, new_width) == (old_height, old_width):
Philip Meier's avatar
Philip Meier committed
334
        return bounding_boxes, canvas_size
335

336
337
    w_ratio = new_width / old_width
    h_ratio = new_height / old_height
338
    ratios = torch.tensor([w_ratio, h_ratio, w_ratio, h_ratio], device=bounding_boxes.device)
339
    return (
340
        bounding_boxes.mul(ratios).to(bounding_boxes.dtype),
341
342
        (new_height, new_width),
    )
343
344


345
346
347
348
349
350
351
@_register_kernel_internal(resize, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _resize_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, size: List[int], max_size: Optional[int] = None, **kwargs: Any
) -> datapoints.BoundingBoxes:
    output, canvas_size = resize_bounding_boxes(
        inpt.as_subclass(torch.Tensor), inpt.canvas_size, size, max_size=max_size
    )
352
    return datapoints.wrap(output, like=inpt, canvas_size=canvas_size)
353
354
355


@_register_kernel_internal(resize, datapoints.Video)
356
357
358
def resize_video(
    video: torch.Tensor,
    size: List[int],
359
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
360
    max_size: Optional[int] = None,
361
    antialias: Optional[Union[str, bool]] = "warn",
362
363
364
365
) -> torch.Tensor:
    return resize_image_tensor(video, size=size, interpolation=interpolation, max_size=max_size, antialias=antialias)


366
def affine(
367
    inpt: torch.Tensor,
368
369
370
371
372
    angle: Union[int, float],
    translate: List[float],
    scale: float,
    shear: List[float],
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
373
    fill: _FillTypeJIT = None,
374
    center: Optional[List[float]] = None,
375
) -> torch.Tensor:
376
    if torch.jit.is_scripting():
377
378
        return affine_image_tensor(
            inpt,
379
            angle=angle,
380
381
382
383
384
385
386
            translate=translate,
            scale=scale,
            shear=shear,
            interpolation=interpolation,
            fill=fill,
            center=center,
        )
387
388
389
390
391
392
393
394
395
396
397
398
399
400

    _log_api_usage_once(affine)

    kernel = _get_kernel(affine, type(inpt))
    return kernel(
        inpt,
        angle=angle,
        translate=translate,
        scale=scale,
        shear=shear,
        interpolation=interpolation,
        fill=fill,
        center=center,
    )
401
402


403
def _affine_parse_args(
404
    angle: Union[int, float],
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
    translate: List[float],
    scale: float,
    shear: List[float],
    interpolation: InterpolationMode = InterpolationMode.NEAREST,
    center: Optional[List[float]] = None,
) -> Tuple[float, List[float], List[float], Optional[List[float]]]:
    if not isinstance(angle, (int, float)):
        raise TypeError("Argument angle should be int or float")

    if not isinstance(translate, (list, tuple)):
        raise TypeError("Argument translate should be a sequence")

    if len(translate) != 2:
        raise ValueError("Argument translate should be a sequence of length 2")

    if scale <= 0.0:
        raise ValueError("Argument scale should be positive")

    if not isinstance(shear, (numbers.Number, (list, tuple))):
        raise TypeError("Shear should be either a single value or a sequence of two values")

    if not isinstance(interpolation, InterpolationMode):
        raise TypeError("Argument interpolation should be a InterpolationMode")

    if isinstance(angle, int):
        angle = float(angle)

    if isinstance(translate, tuple):
        translate = list(translate)

    if isinstance(shear, numbers.Number):
        shear = [shear, 0.0]

    if isinstance(shear, tuple):
        shear = list(shear)

    if len(shear) == 1:
        shear = [shear[0], shear[0]]

    if len(shear) != 2:
        raise ValueError(f"Shear should be a sequence containing two values. Got {shear}")

447
448
449
450
451
    if center is not None:
        if not isinstance(center, (list, tuple)):
            raise TypeError("Argument center should be a sequence")
        else:
            center = [float(c) for c in center]
452
453
454
455

    return angle, translate, shear, center


456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
def _get_inverse_affine_matrix(
    center: List[float], angle: float, translate: List[float], scale: float, shear: List[float], inverted: bool = True
) -> List[float]:
    # Helper method to compute inverse matrix for affine transformation

    # Pillow requires inverse affine transformation matrix:
    # Affine matrix is : M = T * C * RotateScaleShear * C^-1
    #
    # where T is translation matrix: [1, 0, tx | 0, 1, ty | 0, 0, 1]
    #       C is translation matrix to keep center: [1, 0, cx | 0, 1, cy | 0, 0, 1]
    #       RotateScaleShear is rotation with scale and shear matrix
    #
    #       RotateScaleShear(a, s, (sx, sy)) =
    #       = R(a) * S(s) * SHy(sy) * SHx(sx)
    #       = [ s*cos(a - sy)/cos(sy), s*(-cos(a - sy)*tan(sx)/cos(sy) - sin(a)), 0 ]
    #         [ s*sin(a - sy)/cos(sy), s*(-sin(a - sy)*tan(sx)/cos(sy) + cos(a)), 0 ]
    #         [ 0                    , 0                                      , 1 ]
    # where R is a rotation matrix, S is a scaling matrix, and SHx and SHy are the shears:
    # SHx(s) = [1, -tan(s)] and SHy(s) = [1      , 0]
    #          [0, 1      ]              [-tan(s), 1]
    #
    # Thus, the inverse is M^-1 = C * RotateScaleShear^-1 * C^-1 * T^-1

    rot = math.radians(angle)
    sx = math.radians(shear[0])
    sy = math.radians(shear[1])

    cx, cy = center
    tx, ty = translate

    # Cached results
    cos_sy = math.cos(sy)
    tan_sx = math.tan(sx)
    rot_minus_sy = rot - sy
    cx_plus_tx = cx + tx
    cy_plus_ty = cy + ty

    # Rotate Scale Shear (RSS) without scaling
    a = math.cos(rot_minus_sy) / cos_sy
    b = -(a * tan_sx + math.sin(rot))
    c = math.sin(rot_minus_sy) / cos_sy
    d = math.cos(rot) - c * tan_sx

    if inverted:
        # Inverted rotation matrix with scale and shear
        # det([[a, b], [c, d]]) == 1, since det(rotation) = 1 and det(shear) = 1
        matrix = [d / scale, -b / scale, 0.0, -c / scale, a / scale, 0.0]
        # Apply inverse of translation and of center translation: RSS^-1 * C^-1 * T^-1
        # and then apply center translation: C * RSS^-1 * C^-1 * T^-1
        matrix[2] += cx - matrix[0] * cx_plus_tx - matrix[1] * cy_plus_ty
        matrix[5] += cy - matrix[3] * cx_plus_tx - matrix[4] * cy_plus_ty
    else:
        matrix = [a * scale, b * scale, 0.0, c * scale, d * scale, 0.0]
        # Apply inverse of center translation: RSS * C^-1
        # and then apply translation and center : T * C * RSS * C^-1
        matrix[2] += cx_plus_tx - matrix[0] * cx - matrix[1] * cy
        matrix[5] += cy_plus_ty - matrix[3] * cx - matrix[4] * cy

    return matrix


def _compute_affine_output_size(matrix: List[float], w: int, h: int) -> Tuple[int, int]:
    # Inspired of PIL implementation:
    # https://github.com/python-pillow/Pillow/blob/11de3318867e4398057373ee9f12dcb33db7335c/src/PIL/Image.py#L2054

    # pts are Top-Left, Top-Right, Bottom-Left, Bottom-Right points.
    # Points are shifted due to affine matrix torch convention about
    # the center point. Center is (0, 0) for image center pivot point (w * 0.5, h * 0.5)
    half_w = 0.5 * w
    half_h = 0.5 * h
    pts = torch.tensor(
        [
            [-half_w, -half_h, 1.0],
            [-half_w, half_h, 1.0],
            [half_w, half_h, 1.0],
            [half_w, -half_h, 1.0],
        ]
    )
    theta = torch.tensor(matrix, dtype=torch.float).view(2, 3)
    new_pts = torch.matmul(pts, theta.T)
    min_vals, max_vals = new_pts.aminmax(dim=0)

    # shift points to [0, w] and [0, h] interval to match PIL results
    halfs = torch.tensor((half_w, half_h))
    min_vals.add_(halfs)
    max_vals.add_(halfs)

    # Truncate precision to 1e-4 to avoid ceil of Xe-15 to 1.0
    tol = 1e-4
    inv_tol = 1.0 / tol
    cmax = max_vals.mul_(inv_tol).trunc_().mul_(tol).ceil_()
    cmin = min_vals.mul_(inv_tol).trunc_().mul_(tol).floor_()
    size = cmax.sub_(cmin)
    return int(size[0]), int(size[1])  # w, h


552
def _apply_grid_transform(img: torch.Tensor, grid: torch.Tensor, mode: str, fill: _FillTypeJIT) -> torch.Tensor:
553

554
555
556
557
    # We are using context knowledge that grid should have float dtype
    fp = img.dtype == grid.dtype
    float_img = img if fp else img.to(grid.dtype)

558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
    shape = float_img.shape
    if shape[0] > 1:
        # Apply same grid to a batch of images
        grid = grid.expand(shape[0], -1, -1, -1)

    # Append a dummy mask for customized fill colors, should be faster than grid_sample() twice
    if fill is not None:
        mask = torch.ones((shape[0], 1, shape[2], shape[3]), dtype=float_img.dtype, device=float_img.device)
        float_img = torch.cat((float_img, mask), dim=1)

    float_img = grid_sample(float_img, grid, mode=mode, padding_mode="zeros", align_corners=False)

    # Fill with required color
    if fill is not None:
        float_img, mask = torch.tensor_split(float_img, indices=(-1,), dim=-3)
        mask = mask.expand_as(float_img)
574
        fill_list = fill if isinstance(fill, (tuple, list)) else [float(fill)]  # type: ignore[arg-type]
575
576
577
578
579
580
581
582
583
        fill_img = torch.tensor(fill_list, dtype=float_img.dtype, device=float_img.device).view(1, -1, 1, 1)
        if mode == "nearest":
            bool_mask = mask < 0.5
            float_img[bool_mask] = fill_img.expand_as(float_img)[bool_mask]
        else:  # 'bilinear'
            # The following is mathematically equivalent to:
            # img * mask + (1.0 - mask) * fill = img * mask - fill * mask + fill = mask * (img - fill) + fill
            float_img = float_img.sub_(fill_img).mul_(mask).add_(fill_img)

584
585
586
    img = float_img.round_().to(img.dtype) if not fp else float_img

    return img
587
588
589
590
591
592


def _assert_grid_transform_inputs(
    image: torch.Tensor,
    matrix: Optional[List[float]],
    interpolation: str,
593
    fill: _FillTypeJIT,
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
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
    supported_interpolation_modes: List[str],
    coeffs: Optional[List[float]] = None,
) -> None:
    if matrix is not None:
        if not isinstance(matrix, list):
            raise TypeError("Argument matrix should be a list")
        elif len(matrix) != 6:
            raise ValueError("Argument matrix should have 6 float values")

    if coeffs is not None and len(coeffs) != 8:
        raise ValueError("Argument coeffs should have 8 float values")

    if fill is not None:
        if isinstance(fill, (tuple, list)):
            length = len(fill)
            num_channels = image.shape[-3]
            if length > 1 and length != num_channels:
                raise ValueError(
                    "The number of elements in 'fill' cannot broadcast to match the number of "
                    f"channels of the image ({length} != {num_channels})"
                )
        elif not isinstance(fill, (int, float)):
            raise ValueError("Argument fill should be either int, float, tuple or list")

    if interpolation not in supported_interpolation_modes:
        raise ValueError(f"Interpolation mode '{interpolation}' is unsupported with Tensor input")


def _affine_grid(
    theta: torch.Tensor,
    w: int,
    h: int,
    ow: int,
    oh: int,
) -> torch.Tensor:
    # https://github.com/pytorch/pytorch/blob/74b65c32be68b15dc7c9e8bb62459efbfbde33d8/aten/src/ATen/native/
    # AffineGridGenerator.cpp#L18
    # Difference with AffineGridGenerator is that:
    # 1) we normalize grid values after applying theta
    # 2) we can normalize by other image size, such that it covers "extend" option like in PIL.Image.rotate
    dtype = theta.dtype
    device = theta.device

    base_grid = torch.empty(1, oh, ow, 3, dtype=dtype, device=device)
    x_grid = torch.linspace((1.0 - ow) * 0.5, (ow - 1.0) * 0.5, steps=ow, device=device)
    base_grid[..., 0].copy_(x_grid)
    y_grid = torch.linspace((1.0 - oh) * 0.5, (oh - 1.0) * 0.5, steps=oh, device=device).unsqueeze_(-1)
    base_grid[..., 1].copy_(y_grid)
    base_grid[..., 2].fill_(1)

    rescaled_theta = theta.transpose(1, 2).div_(torch.tensor([0.5 * w, 0.5 * h], dtype=dtype, device=device))
    output_grid = base_grid.view(1, oh * ow, 3).bmm(rescaled_theta)
    return output_grid.view(1, oh, ow, 2)


649
@_register_kernel_internal(affine, torch.Tensor)
650
@_register_kernel_internal(affine, datapoints.Image)
651
def affine_image_tensor(
652
    image: torch.Tensor,
653
    angle: Union[int, float],
654
655
656
    translate: List[float],
    scale: float,
    shear: List[float],
657
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
658
    fill: _FillTypeJIT = None,
659
660
    center: Optional[List[float]] = None,
) -> torch.Tensor:
661
662
    interpolation = _check_interpolation(interpolation)

663
664
    if image.numel() == 0:
        return image
665

666
    shape = image.shape
667
    ndim = image.ndim
668

669
670
671
672
673
674
675
676
677
678
    if ndim > 4:
        image = image.reshape((-1,) + shape[-3:])
        needs_unsquash = True
    elif ndim == 3:
        image = image.unsqueeze(0)
        needs_unsquash = True
    else:
        needs_unsquash = False

    height, width = shape[-2:]
679
680
681
682
683
    angle, translate, shear, center = _affine_parse_args(angle, translate, scale, shear, interpolation, center)

    center_f = [0.0, 0.0]
    if center is not None:
        # Center values should be in pixel coordinates but translated such that (0, 0) corresponds to image center.
684
        center_f = [(c - s * 0.5) for c, s in zip(center, [width, height])]
685

686
    translate_f = [float(t) for t in translate]
687
688
    matrix = _get_inverse_affine_matrix(center_f, angle, translate_f, scale, shear)

689
690
    _assert_grid_transform_inputs(image, matrix, interpolation.value, fill, ["nearest", "bilinear"])

691
    dtype = image.dtype if torch.is_floating_point(image) else torch.float32
692
693
    theta = torch.tensor(matrix, dtype=dtype, device=image.device).reshape(1, 2, 3)
    grid = _affine_grid(theta, w=width, h=height, ow=width, oh=height)
694
    output = _apply_grid_transform(image, grid, interpolation.value, fill=fill)
695
696
697
698
699

    if needs_unsquash:
        output = output.reshape(shape)

    return output
700
701


702
@_register_kernel_internal(affine, PIL.Image.Image)
703
def affine_image_pil(
704
    image: PIL.Image.Image,
705
    angle: Union[int, float],
706
707
708
    translate: List[float],
    scale: float,
    shear: List[float],
709
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
710
    fill: _FillTypeJIT = None,
711
712
    center: Optional[List[float]] = None,
) -> PIL.Image.Image:
713
    interpolation = _check_interpolation(interpolation)
714
715
716
717
718
719
    angle, translate, shear, center = _affine_parse_args(angle, translate, scale, shear, interpolation, center)

    # center = (img_size[0] * 0.5 + 0.5, img_size[1] * 0.5 + 0.5)
    # it is visually better to estimate the center without 0.5 offset
    # otherwise image rotated by 90 degrees is shifted vs output image of torch.rot90 or F_t.affine
    if center is None:
Philip Meier's avatar
Philip Meier committed
720
        height, width = get_size_image_pil(image)
721
722
723
        center = [width * 0.5, height * 0.5]
    matrix = _get_inverse_affine_matrix(center, angle, translate, scale, shear)

724
    return _FP.affine(image, matrix, interpolation=pil_modes_mapping[interpolation], fill=fill)
725
726


727
728
def _affine_bounding_boxes_with_expand(
    bounding_boxes: torch.Tensor,
729
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
730
    canvas_size: Tuple[int, int],
731
732
733
734
    angle: Union[int, float],
    translate: List[float],
    scale: float,
    shear: List[float],
735
    center: Optional[List[float]] = None,
736
    expand: bool = False,
737
) -> Tuple[torch.Tensor, Tuple[int, int]]:
738
    if bounding_boxes.numel() == 0:
Philip Meier's avatar
Philip Meier committed
739
        return bounding_boxes, canvas_size
740
741
742
743
744
745
746
747
748

    original_shape = bounding_boxes.shape
    original_dtype = bounding_boxes.dtype
    bounding_boxes = bounding_boxes.clone() if bounding_boxes.is_floating_point() else bounding_boxes.float()
    dtype = bounding_boxes.dtype
    device = bounding_boxes.device
    bounding_boxes = (
        convert_format_bounding_boxes(
            bounding_boxes, old_format=format, new_format=datapoints.BoundingBoxFormat.XYXY, inplace=True
749
750
751
        )
    ).reshape(-1, 4)

752
753
754
    angle, translate, shear, center = _affine_parse_args(
        angle, translate, scale, shear, InterpolationMode.NEAREST, center
    )
755

756
    if center is None:
Philip Meier's avatar
Philip Meier committed
757
        height, width = canvas_size
758
759
        center = [width * 0.5, height * 0.5]

760
761
762
763
764
765
766
    affine_vector = _get_inverse_affine_matrix(center, angle, translate, scale, shear, inverted=False)
    transposed_affine_matrix = (
        torch.tensor(
            affine_vector,
            dtype=dtype,
            device=device,
        )
767
        .reshape(2, 3)
768
769
        .T
    )
770
771
772
773
    # 1) Let's transform bboxes into a tensor of 4 points (top-left, top-right, bottom-left, bottom-right corners).
    # Tensor of points has shape (N * 4, 3), where N is the number of bboxes
    # Single point structure is similar to
    # [(xmin, ymin, 1), (xmax, ymin, 1), (xmax, ymax, 1), (xmin, ymax, 1)]
774
    points = bounding_boxes[:, [[0, 1], [2, 1], [2, 3], [0, 3]]].reshape(-1, 2)
775
    points = torch.cat([points, torch.ones(points.shape[0], 1, device=device, dtype=dtype)], dim=-1)
776
    # 2) Now let's transform the points using affine matrix
777
    transformed_points = torch.matmul(points, transposed_affine_matrix)
778
779
    # 3) Reshape transformed points to [N boxes, 4 points, x/y coords]
    # and compute bounding box from 4 transformed points:
780
    transformed_points = transformed_points.reshape(-1, 4, 2)
781
    out_bbox_mins, out_bbox_maxs = torch.aminmax(transformed_points, dim=1)
782
    out_bboxes = torch.cat([out_bbox_mins, out_bbox_maxs], dim=1)
783
784
785
786

    if expand:
        # Compute minimum point for transformed image frame:
        # Points are Top-Left, Top-Right, Bottom-Left, Bottom-Right points.
Philip Meier's avatar
Philip Meier committed
787
        height, width = canvas_size
788
789
790
        points = torch.tensor(
            [
                [0.0, 0.0, 1.0],
791
792
793
                [0.0, float(height), 1.0],
                [float(width), float(height), 1.0],
                [float(width), 0.0, 1.0],
794
795
796
797
            ],
            dtype=dtype,
            device=device,
        )
798
        new_points = torch.matmul(points, transposed_affine_matrix)
799
        tr = torch.amin(new_points, dim=0, keepdim=True)
800
        # Translate bounding boxes
801
        out_bboxes.sub_(tr.repeat((1, 2)))
802
803
        # Estimate meta-data for image with inverted=True and with center=[0,0]
        affine_vector = _get_inverse_affine_matrix([0.0, 0.0], angle, translate, scale, shear)
804
        new_width, new_height = _compute_affine_output_size(affine_vector, width, height)
Philip Meier's avatar
Philip Meier committed
805
        canvas_size = (new_height, new_width)
806

Philip Meier's avatar
Philip Meier committed
807
    out_bboxes = clamp_bounding_boxes(out_bboxes, format=datapoints.BoundingBoxFormat.XYXY, canvas_size=canvas_size)
808
    out_bboxes = convert_format_bounding_boxes(
809
810
811
812
        out_bboxes, old_format=datapoints.BoundingBoxFormat.XYXY, new_format=format, inplace=True
    ).reshape(original_shape)

    out_bboxes = out_bboxes.to(original_dtype)
Philip Meier's avatar
Philip Meier committed
813
    return out_bboxes, canvas_size
814
815


816
817
def affine_bounding_boxes(
    bounding_boxes: torch.Tensor,
818
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
819
    canvas_size: Tuple[int, int],
820
    angle: Union[int, float],
821
822
823
824
825
    translate: List[float],
    scale: float,
    shear: List[float],
    center: Optional[List[float]] = None,
) -> torch.Tensor:
826
827
    out_box, _ = _affine_bounding_boxes_with_expand(
        bounding_boxes,
828
        format=format,
Philip Meier's avatar
Philip Meier committed
829
        canvas_size=canvas_size,
830
831
832
833
834
835
836
837
        angle=angle,
        translate=translate,
        scale=scale,
        shear=shear,
        center=center,
        expand=False,
    )
    return out_box
838
839


840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
@_register_kernel_internal(affine, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _affine_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes,
    angle: Union[int, float],
    translate: List[float],
    scale: float,
    shear: List[float],
    center: Optional[List[float]] = None,
    **kwargs,
) -> datapoints.BoundingBoxes:
    output = affine_bounding_boxes(
        inpt.as_subclass(torch.Tensor),
        format=inpt.format,
        canvas_size=inpt.canvas_size,
        angle=angle,
        translate=translate,
        scale=scale,
        shear=shear,
        center=center,
    )
860
    return datapoints.wrap(output, like=inpt)
861
862


863
864
def affine_mask(
    mask: torch.Tensor,
865
    angle: Union[int, float],
866
867
868
    translate: List[float],
    scale: float,
    shear: List[float],
869
    fill: _FillTypeJIT = None,
870
871
    center: Optional[List[float]] = None,
) -> torch.Tensor:
872
873
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
874
875
876
877
878
        needs_squeeze = True
    else:
        needs_squeeze = False

    output = affine_image_tensor(
879
        mask,
880
881
882
883
884
        angle=angle,
        translate=translate,
        scale=scale,
        shear=shear,
        interpolation=InterpolationMode.NEAREST,
885
        fill=fill,
886
887
888
        center=center,
    )

889
890
891
892
893
    if needs_squeeze:
        output = output.squeeze(0)

    return output

894

895
896
897
898
899
900
901
@_register_kernel_internal(affine, datapoints.Mask, datapoint_wrapper=False)
def _affine_mask_dispatch(
    inpt: datapoints.Mask,
    angle: Union[int, float],
    translate: List[float],
    scale: float,
    shear: List[float],
902
    fill: _FillTypeJIT = None,
903
904
905
906
907
908
909
910
911
912
913
914
    center: Optional[List[float]] = None,
    **kwargs,
) -> datapoints.Mask:
    output = affine_mask(
        inpt.as_subclass(torch.Tensor),
        angle=angle,
        translate=translate,
        scale=scale,
        shear=shear,
        fill=fill,
        center=center,
    )
915
    return datapoints.wrap(output, like=inpt)
916
917
918


@_register_kernel_internal(affine, datapoints.Video)
919
920
921
922
923
924
def affine_video(
    video: torch.Tensor,
    angle: Union[int, float],
    translate: List[float],
    scale: float,
    shear: List[float],
925
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
926
    fill: _FillTypeJIT = None,
927
928
929
930
931
932
933
934
935
936
937
938
939
940
    center: Optional[List[float]] = None,
) -> torch.Tensor:
    return affine_image_tensor(
        video,
        angle=angle,
        translate=translate,
        scale=scale,
        shear=shear,
        interpolation=interpolation,
        fill=fill,
        center=center,
    )


941
def rotate(
942
    inpt: torch.Tensor,
943
    angle: float,
944
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
945
    expand: bool = False,
946
    center: Optional[List[float]] = None,
947
948
    fill: _FillTypeJIT = None,
) -> torch.Tensor:
949
950
951
    if torch.jit.is_scripting():
        return rotate_image_tensor(
            inpt, angle=angle, interpolation=interpolation, expand=expand, fill=fill, center=center
952
        )
953

954
    _log_api_usage_once(rotate)
955

956
957
958
959
960
    kernel = _get_kernel(rotate, type(inpt))
    return kernel(inpt, angle=angle, interpolation=interpolation, expand=expand, fill=fill, center=center)


@_register_kernel_internal(rotate, torch.Tensor)
961
@_register_kernel_internal(rotate, datapoints.Image)
962
def rotate_image_tensor(
963
    image: torch.Tensor,
964
    angle: float,
965
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
966
967
    expand: bool = False,
    center: Optional[List[float]] = None,
968
    fill: _FillTypeJIT = None,
969
) -> torch.Tensor:
970
971
    interpolation = _check_interpolation(interpolation)

972
973
    shape = image.shape
    num_channels, height, width = shape[-3:]
974

975
976
    center_f = [0.0, 0.0]
    if center is not None:
977
        if expand:
978
            # TODO: Do we actually want to warn, or just document this?
979
            warnings.warn("The provided center argument has no effect on the result if expand is True")
980
981
        # Center values should be in pixel coordinates but translated such that (0, 0) corresponds to image center.
        center_f = [(c - s * 0.5) for c, s in zip(center, [width, height])]
982
983
984
985

    # due to current incoherence of rotation angle direction between affine and rotate implementations
    # we need to set -angle.
    matrix = _get_inverse_affine_matrix(center_f, -angle, [0.0, 0.0], 1.0, [0.0, 0.0])
986

987
    if image.numel() > 0:
988
989
990
991
992
        image = image.reshape(-1, num_channels, height, width)

        _assert_grid_transform_inputs(image, matrix, interpolation.value, fill, ["nearest", "bilinear"])

        ow, oh = _compute_affine_output_size(matrix, width, height) if expand else (width, height)
993
        dtype = image.dtype if torch.is_floating_point(image) else torch.float32
994
995
        theta = torch.tensor(matrix, dtype=dtype, device=image.device).reshape(1, 2, 3)
        grid = _affine_grid(theta, w=width, h=height, ow=ow, oh=oh)
996
        output = _apply_grid_transform(image, grid, interpolation.value, fill=fill)
997
998

        new_height, new_width = output.shape[-2:]
999
    else:
1000
1001
        output = image
        new_width, new_height = _compute_affine_output_size(matrix, width, height) if expand else (width, height)
1002

1003
    return output.reshape(shape[:-3] + (num_channels, new_height, new_width))
1004
1005


1006
@_register_kernel_internal(rotate, PIL.Image.Image)
1007
def rotate_image_pil(
1008
    image: PIL.Image.Image,
1009
    angle: float,
1010
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
1011
1012
    expand: bool = False,
    center: Optional[List[float]] = None,
1013
    fill: _FillTypeJIT = None,
1014
) -> PIL.Image.Image:
1015
1016
    interpolation = _check_interpolation(interpolation)

1017
    if center is not None and expand:
1018
        warnings.warn("The provided center argument has no effect on the result if expand is True")
1019

1020
    return _FP.rotate(
1021
        image, angle, interpolation=pil_modes_mapping[interpolation], expand=expand, fill=fill, center=center
1022
1023
1024
    )


1025
1026
def rotate_bounding_boxes(
    bounding_boxes: torch.Tensor,
1027
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
1028
    canvas_size: Tuple[int, int],
1029
1030
1031
    angle: float,
    expand: bool = False,
    center: Optional[List[float]] = None,
1032
) -> Tuple[torch.Tensor, Tuple[int, int]]:
1033
1034
1035
    if center is not None and expand:
        warnings.warn("The provided center argument has no effect on the result if expand is True")

1036
1037
    return _affine_bounding_boxes_with_expand(
        bounding_boxes,
1038
        format=format,
Philip Meier's avatar
Philip Meier committed
1039
        canvas_size=canvas_size,
1040
1041
1042
1043
1044
1045
1046
        angle=-angle,
        translate=[0.0, 0.0],
        scale=1.0,
        shear=[0.0, 0.0],
        center=center,
        expand=expand,
    )
1047
1048


1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
@_register_kernel_internal(rotate, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _rotate_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, angle: float, expand: bool = False, center: Optional[List[float]] = None, **kwargs
) -> datapoints.BoundingBoxes:
    output, canvas_size = rotate_bounding_boxes(
        inpt.as_subclass(torch.Tensor),
        format=inpt.format,
        canvas_size=inpt.canvas_size,
        angle=angle,
        expand=expand,
        center=center,
    )
1061
    return datapoints.wrap(output, like=inpt, canvas_size=canvas_size)
1062
1063


1064
1065
def rotate_mask(
    mask: torch.Tensor,
1066
1067
1068
    angle: float,
    expand: bool = False,
    center: Optional[List[float]] = None,
1069
    fill: _FillTypeJIT = None,
1070
) -> torch.Tensor:
1071
1072
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
1073
1074
1075
1076
1077
        needs_squeeze = True
    else:
        needs_squeeze = False

    output = rotate_image_tensor(
1078
        mask,
1079
1080
1081
        angle=angle,
        expand=expand,
        interpolation=InterpolationMode.NEAREST,
1082
        fill=fill,
1083
1084
1085
        center=center,
    )

1086
1087
1088
1089
1090
    if needs_squeeze:
        output = output.squeeze(0)

    return output

1091

1092
1093
1094
1095
1096
1097
@_register_kernel_internal(rotate, datapoints.Mask, datapoint_wrapper=False)
def _rotate_mask_dispatch(
    inpt: datapoints.Mask,
    angle: float,
    expand: bool = False,
    center: Optional[List[float]] = None,
1098
    fill: _FillTypeJIT = None,
1099
1100
1101
    **kwargs,
) -> datapoints.Mask:
    output = rotate_mask(inpt.as_subclass(torch.Tensor), angle=angle, expand=expand, fill=fill, center=center)
1102
    return datapoints.wrap(output, like=inpt)
1103
1104
1105


@_register_kernel_internal(rotate, datapoints.Video)
1106
1107
1108
def rotate_video(
    video: torch.Tensor,
    angle: float,
1109
    interpolation: Union[InterpolationMode, int] = InterpolationMode.NEAREST,
1110
1111
    expand: bool = False,
    center: Optional[List[float]] = None,
1112
    fill: _FillTypeJIT = None,
1113
1114
1115
1116
) -> torch.Tensor:
    return rotate_image_tensor(video, angle, interpolation=interpolation, expand=expand, fill=fill, center=center)


1117
def pad(
1118
    inpt: torch.Tensor,
1119
1120
1121
    padding: List[int],
    fill: Optional[Union[int, float, List[float]]] = None,
    padding_mode: str = "constant",
1122
) -> torch.Tensor:
1123
1124
    if torch.jit.is_scripting():
        return pad_image_tensor(inpt, padding=padding, fill=fill, padding_mode=padding_mode)
1125

1126
    _log_api_usage_once(pad)
1127

1128
1129
    kernel = _get_kernel(pad, type(inpt))
    return kernel(inpt, padding=padding, fill=fill, padding_mode=padding_mode)
1130
1131


1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
def _parse_pad_padding(padding: Union[int, List[int]]) -> List[int]:
    if isinstance(padding, int):
        pad_left = pad_right = pad_top = pad_bottom = padding
    elif isinstance(padding, (tuple, list)):
        if len(padding) == 1:
            pad_left = pad_right = pad_top = pad_bottom = padding[0]
        elif len(padding) == 2:
            pad_left = pad_right = padding[0]
            pad_top = pad_bottom = padding[1]
        elif len(padding) == 4:
            pad_left = padding[0]
            pad_top = padding[1]
            pad_right = padding[2]
            pad_bottom = padding[3]
        else:
            raise ValueError(
                f"Padding must be an int or a 1, 2, or 4 element tuple, not a {len(padding)} element tuple"
            )
    else:
        raise TypeError(f"`padding` should be an integer or tuple or list of integers, but got {padding}")

    return [pad_left, pad_right, pad_top, pad_bottom]
1154

1155

1156
@_register_kernel_internal(pad, torch.Tensor)
1157
@_register_kernel_internal(pad, datapoints.Image)
1158
def pad_image_tensor(
1159
    image: torch.Tensor,
1160
1161
    padding: List[int],
    fill: Optional[Union[int, float, List[float]]] = None,
1162
1163
    padding_mode: str = "constant",
) -> torch.Tensor:
1164
1165
1166
1167
1168
    # Be aware that while `padding` has order `[left, top, right, bottom]` has order, `torch_padding` uses
    # `[left, right, top, bottom]`. This stems from the fact that we align our API with PIL, but need to use `torch_pad`
    # internally.
    torch_padding = _parse_pad_padding(padding)

1169
    if padding_mode not in ("constant", "edge", "reflect", "symmetric"):
1170
1171
1172
1173
1174
        raise ValueError(
            f"`padding_mode` should be either `'constant'`, `'edge'`, `'reflect'` or `'symmetric'`, "
            f"but got `'{padding_mode}'`."
        )

1175
    if fill is None:
1176
1177
1178
1179
1180
1181
        fill = 0

    if isinstance(fill, (int, float)):
        return _pad_with_scalar_fill(image, torch_padding, fill=fill, padding_mode=padding_mode)
    elif len(fill) == 1:
        return _pad_with_scalar_fill(image, torch_padding, fill=fill[0], padding_mode=padding_mode)
1182
    else:
1183
        return _pad_with_vector_fill(image, torch_padding, fill=fill, padding_mode=padding_mode)
1184
1185
1186


def _pad_with_scalar_fill(
1187
    image: torch.Tensor,
1188
1189
1190
    torch_padding: List[int],
    fill: Union[int, float],
    padding_mode: str,
1191
) -> torch.Tensor:
1192
1193
    shape = image.shape
    num_channels, height, width = shape[-3:]
1194

1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
    batch_size = 1
    for s in shape[:-3]:
        batch_size *= s

    image = image.reshape(batch_size, num_channels, height, width)

    if padding_mode == "edge":
        # Similar to the padding order, `torch_pad`'s PIL's padding modes don't have the same names. Thus, we map
        # the PIL name for the padding mode, which we are also using for our API, to the corresponding `torch_pad`
        # name.
        padding_mode = "replicate"

    if padding_mode == "constant":
        image = torch_pad(image, torch_padding, mode=padding_mode, value=float(fill))
    elif padding_mode in ("reflect", "replicate"):
        # `torch_pad` only supports `"reflect"` or `"replicate"` padding for floating point inputs.
        # TODO: See https://github.com/pytorch/pytorch/issues/40763
        dtype = image.dtype
        if not image.is_floating_point():
            needs_cast = True
            image = image.to(torch.float32)
        else:
            needs_cast = False
1218

1219
1220
1221
1222
1223
        image = torch_pad(image, torch_padding, mode=padding_mode)

        if needs_cast:
            image = image.to(dtype)
    else:  # padding_mode == "symmetric"
1224
        image = _pad_symmetric(image, torch_padding)
1225
1226

    new_height, new_width = image.shape[-2:]
1227

1228
    return image.reshape(shape[:-3] + (num_channels, new_height, new_width))
1229
1230


1231
# TODO: This should be removed once torch_pad supports non-scalar padding values
1232
def _pad_with_vector_fill(
1233
    image: torch.Tensor,
1234
    torch_padding: List[int],
1235
    fill: List[float],
1236
    padding_mode: str,
1237
1238
1239
1240
) -> torch.Tensor:
    if padding_mode != "constant":
        raise ValueError(f"Padding mode '{padding_mode}' is not supported if fill is not scalar")

1241
1242
    output = _pad_with_scalar_fill(image, torch_padding, fill=0, padding_mode="constant")
    left, right, top, bottom = torch_padding
1243
    fill = torch.tensor(fill, dtype=image.dtype, device=image.device).reshape(-1, 1, 1)
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255

    if top > 0:
        output[..., :top, :] = fill
    if left > 0:
        output[..., :, :left] = fill
    if bottom > 0:
        output[..., -bottom:, :] = fill
    if right > 0:
        output[..., :, -right:] = fill
    return output


1256
pad_image_pil = _register_kernel_internal(pad, PIL.Image.Image)(_FP.pad)
1257
1258


1259
@_register_kernel_internal(pad, datapoints.Mask)
1260
1261
def pad_mask(
    mask: torch.Tensor,
1262
1263
    padding: List[int],
    fill: Optional[Union[int, float, List[float]]] = None,
1264
1265
    padding_mode: str = "constant",
) -> torch.Tensor:
1266
1267
1268
    if fill is None:
        fill = 0

1269
    if isinstance(fill, (tuple, list)):
1270
1271
        raise ValueError("Non-scalar fill value is not supported")

1272
1273
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
1274
1275
1276
1277
        needs_squeeze = True
    else:
        needs_squeeze = False

1278
    output = pad_image_tensor(mask, padding=padding, fill=fill, padding_mode=padding_mode)
1279
1280
1281
1282
1283

    if needs_squeeze:
        output = output.squeeze(0)

    return output
1284
1285


1286
1287
def pad_bounding_boxes(
    bounding_boxes: torch.Tensor,
1288
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
1289
    canvas_size: Tuple[int, int],
1290
    padding: List[int],
vfdev's avatar
vfdev committed
1291
    padding_mode: str = "constant",
1292
) -> Tuple[torch.Tensor, Tuple[int, int]]:
vfdev's avatar
vfdev committed
1293
1294
1295
1296
    if padding_mode not in ["constant"]:
        # TODO: add support of other padding modes
        raise ValueError(f"Padding mode '{padding_mode}' is not supported with bounding boxes")

1297
    left, right, top, bottom = _parse_pad_padding(padding)
1298

1299
    if format == datapoints.BoundingBoxFormat.XYXY:
1300
1301
1302
        pad = [left, top, left, top]
    else:
        pad = [left, top, 0, 0]
1303
    bounding_boxes = bounding_boxes + torch.tensor(pad, dtype=bounding_boxes.dtype, device=bounding_boxes.device)
1304

Philip Meier's avatar
Philip Meier committed
1305
    height, width = canvas_size
1306
1307
    height += top + bottom
    width += left + right
Philip Meier's avatar
Philip Meier committed
1308
    canvas_size = (height, width)
1309

Philip Meier's avatar
Philip Meier committed
1310
    return clamp_bounding_boxes(bounding_boxes, format=format, canvas_size=canvas_size), canvas_size
1311
1312


1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
@_register_kernel_internal(pad, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _pad_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, padding: List[int], padding_mode: str = "constant", **kwargs
) -> datapoints.BoundingBoxes:
    output, canvas_size = pad_bounding_boxes(
        inpt.as_subclass(torch.Tensor),
        format=inpt.format,
        canvas_size=inpt.canvas_size,
        padding=padding,
        padding_mode=padding_mode,
    )
1324
    return datapoints.wrap(output, like=inpt, canvas_size=canvas_size)
1325
1326
1327


@_register_kernel_internal(pad, datapoints.Video)
1328
1329
def pad_video(
    video: torch.Tensor,
1330
1331
    padding: List[int],
    fill: Optional[Union[int, float, List[float]]] = None,
1332
1333
1334
1335
1336
    padding_mode: str = "constant",
) -> torch.Tensor:
    return pad_image_tensor(video, padding, fill=fill, padding_mode=padding_mode)


1337
def crop(inpt: torch.Tensor, top: int, left: int, height: int, width: int) -> torch.Tensor:
1338
1339
1340
1341
    if torch.jit.is_scripting():
        return crop_image_tensor(inpt, top=top, left=left, height=height, width=width)

    _log_api_usage_once(crop)
1342

1343
1344
    kernel = _get_kernel(crop, type(inpt))
    return kernel(inpt, top=top, left=left, height=height, width=width)
1345

1346
1347

@_register_kernel_internal(crop, torch.Tensor)
1348
@_register_kernel_internal(crop, datapoints.Image)
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
def crop_image_tensor(image: torch.Tensor, top: int, left: int, height: int, width: int) -> torch.Tensor:
    h, w = image.shape[-2:]

    right = left + width
    bottom = top + height

    if left < 0 or top < 0 or right > w or bottom > h:
        image = image[..., max(top, 0) : bottom, max(left, 0) : right]
        torch_padding = [
            max(min(right, 0) - left, 0),
            max(right - max(w, left), 0),
            max(min(bottom, 0) - top, 0),
            max(bottom - max(h, top), 0),
        ]
        return _pad_with_scalar_fill(image, torch_padding, fill=0, padding_mode="constant")
    return image[..., top:bottom, left:right]


1367
crop_image_pil = _FP.crop
1368
_register_kernel_internal(crop, PIL.Image.Image)(crop_image_pil)
1369
1370


1371
1372
def crop_bounding_boxes(
    bounding_boxes: torch.Tensor,
1373
    format: datapoints.BoundingBoxFormat,
1374
1375
    top: int,
    left: int,
1376
1377
1378
    height: int,
    width: int,
) -> Tuple[torch.Tensor, Tuple[int, int]]:
1379

1380
    # Crop or implicit pad if left and/or top have negative values:
1381
    if format == datapoints.BoundingBoxFormat.XYXY:
1382
        sub = [left, top, left, top]
1383
    else:
1384
1385
        sub = [left, top, 0, 0]

1386
    bounding_boxes = bounding_boxes - torch.tensor(sub, dtype=bounding_boxes.dtype, device=bounding_boxes.device)
Philip Meier's avatar
Philip Meier committed
1387
    canvas_size = (height, width)
1388

Philip Meier's avatar
Philip Meier committed
1389
    return clamp_bounding_boxes(bounding_boxes, format=format, canvas_size=canvas_size), canvas_size
1390
1391


1392
1393
1394
1395
1396
1397
1398
@_register_kernel_internal(crop, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _crop_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, top: int, left: int, height: int, width: int
) -> datapoints.BoundingBoxes:
    output, canvas_size = crop_bounding_boxes(
        inpt.as_subclass(torch.Tensor), format=inpt.format, top=top, left=left, height=height, width=width
    )
1399
    return datapoints.wrap(output, like=inpt, canvas_size=canvas_size)
1400
1401
1402


@_register_kernel_internal(crop, datapoints.Mask)
1403
def crop_mask(mask: torch.Tensor, top: int, left: int, height: int, width: int) -> torch.Tensor:
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
        needs_squeeze = True
    else:
        needs_squeeze = False

    output = crop_image_tensor(mask, top, left, height, width)

    if needs_squeeze:
        output = output.squeeze(0)

    return output
1416
1417


1418
@_register_kernel_internal(crop, datapoints.Video)
1419
1420
1421
1422
def crop_video(video: torch.Tensor, top: int, left: int, height: int, width: int) -> torch.Tensor:
    return crop_image_tensor(video, top, left, height, width)


1423
def perspective(
1424
    inpt: torch.Tensor,
1425
1426
1427
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1428
    fill: _FillTypeJIT = None,
1429
    coefficients: Optional[List[float]] = None,
1430
) -> torch.Tensor:
1431
    if torch.jit.is_scripting():
1432
        return perspective_image_tensor(
1433
1434
1435
1436
1437
1438
            inpt,
            startpoints=startpoints,
            endpoints=endpoints,
            interpolation=interpolation,
            fill=fill,
            coefficients=coefficients,
1439
        )
1440

1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
    _log_api_usage_once(perspective)

    kernel = _get_kernel(perspective, type(inpt))
    return kernel(
        inpt,
        startpoints=startpoints,
        endpoints=endpoints,
        interpolation=interpolation,
        fill=fill,
        coefficients=coefficients,
    )

1453

1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
def _perspective_grid(coeffs: List[float], ow: int, oh: int, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
    # https://github.com/python-pillow/Pillow/blob/4634eafe3c695a014267eefdce830b4a825beed7/
    # src/libImaging/Geometry.c#L394

    #
    # x_out = (coeffs[0] * x + coeffs[1] * y + coeffs[2]) / (coeffs[6] * x + coeffs[7] * y + 1)
    # y_out = (coeffs[3] * x + coeffs[4] * y + coeffs[5]) / (coeffs[6] * x + coeffs[7] * y + 1)
    #
    theta1 = torch.tensor(
        [[[coeffs[0], coeffs[1], coeffs[2]], [coeffs[3], coeffs[4], coeffs[5]]]], dtype=dtype, device=device
    )
    theta2 = torch.tensor([[[coeffs[6], coeffs[7], 1.0], [coeffs[6], coeffs[7], 1.0]]], dtype=dtype, device=device)

    d = 0.5
    base_grid = torch.empty(1, oh, ow, 3, dtype=dtype, device=device)
1469
    x_grid = torch.linspace(d, ow + d - 1.0, steps=ow, device=device, dtype=dtype)
1470
    base_grid[..., 0].copy_(x_grid)
1471
    y_grid = torch.linspace(d, oh + d - 1.0, steps=oh, device=device, dtype=dtype).unsqueeze_(-1)
1472
1473
1474
1475
    base_grid[..., 1].copy_(y_grid)
    base_grid[..., 2].fill_(1)

    rescaled_theta1 = theta1.transpose(1, 2).div_(torch.tensor([0.5 * ow, 0.5 * oh], dtype=dtype, device=device))
1476
1477
1478
    shape = (1, oh * ow, 3)
    output_grid1 = base_grid.view(shape).bmm(rescaled_theta1)
    output_grid2 = base_grid.view(shape).bmm(theta2.transpose(1, 2))
1479
1480
1481
1482
1483

    output_grid = output_grid1.div_(output_grid2).sub_(1.0)
    return output_grid.view(1, oh, ow, 2)


1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
def _perspective_coefficients(
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
    coefficients: Optional[List[float]],
) -> List[float]:
    if coefficients is not None:
        if startpoints is not None and endpoints is not None:
            raise ValueError("The startpoints/endpoints and the coefficients shouldn't be defined concurrently.")
        elif len(coefficients) != 8:
            raise ValueError("Argument coefficients should have 8 float values")
        return coefficients
    elif startpoints is not None and endpoints is not None:
        return _get_perspective_coeffs(startpoints, endpoints)
    else:
        raise ValueError("Either the startpoints/endpoints or the coefficients must have non `None` values.")


1501
@_register_kernel_internal(perspective, torch.Tensor)
1502
@_register_kernel_internal(perspective, datapoints.Image)
1503
def perspective_image_tensor(
1504
    image: torch.Tensor,
1505
1506
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
1507
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1508
    fill: _FillTypeJIT = None,
1509
    coefficients: Optional[List[float]] = None,
1510
) -> torch.Tensor:
1511
    perspective_coeffs = _perspective_coefficients(startpoints, endpoints, coefficients)
1512
1513
    interpolation = _check_interpolation(interpolation)

1514
1515
1516
1517
    if image.numel() == 0:
        return image

    shape = image.shape
1518
    ndim = image.ndim
1519

1520
    if ndim > 4:
1521
        image = image.reshape((-1,) + shape[-3:])
1522
        needs_unsquash = True
1523
1524
1525
    elif ndim == 3:
        image = image.unsqueeze(0)
        needs_unsquash = True
1526
1527
1528
    else:
        needs_unsquash = False

1529
    _assert_grid_transform_inputs(
1530
1531
1532
1533
1534
1535
1536
1537
        image,
        matrix=None,
        interpolation=interpolation.value,
        fill=fill,
        supported_interpolation_modes=["nearest", "bilinear"],
        coeffs=perspective_coeffs,
    )

1538
    oh, ow = shape[-2:]
1539
    dtype = image.dtype if torch.is_floating_point(image) else torch.float32
1540
    grid = _perspective_grid(perspective_coeffs, ow=ow, oh=oh, dtype=dtype, device=image.device)
1541
    output = _apply_grid_transform(image, grid, interpolation.value, fill=fill)
1542
1543

    if needs_unsquash:
1544
        output = output.reshape(shape)
1545
1546

    return output
1547
1548


1549
@_register_kernel_internal(perspective, PIL.Image.Image)
1550
def perspective_image_pil(
1551
    image: PIL.Image.Image,
1552
1553
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
1554
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BICUBIC,
1555
    fill: _FillTypeJIT = None,
1556
    coefficients: Optional[List[float]] = None,
1557
) -> PIL.Image.Image:
1558
    perspective_coeffs = _perspective_coefficients(startpoints, endpoints, coefficients)
1559
    interpolation = _check_interpolation(interpolation)
1560
    return _FP.perspective(image, perspective_coeffs, interpolation=pil_modes_mapping[interpolation], fill=fill)
1561
1562


1563
1564
def perspective_bounding_boxes(
    bounding_boxes: torch.Tensor,
1565
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
1566
    canvas_size: Tuple[int, int],
1567
1568
1569
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
    coefficients: Optional[List[float]] = None,
1570
) -> torch.Tensor:
1571
1572
    if bounding_boxes.numel() == 0:
        return bounding_boxes
1573

1574
    perspective_coeffs = _perspective_coefficients(startpoints, endpoints, coefficients)
1575

1576
1577
1578
1579
    original_shape = bounding_boxes.shape
    # TODO: first cast to float if bbox is int64 before convert_format_bounding_boxes
    bounding_boxes = (
        convert_format_bounding_boxes(bounding_boxes, old_format=format, new_format=datapoints.BoundingBoxFormat.XYXY)
1580
    ).reshape(-1, 4)
1581

1582
1583
    dtype = bounding_boxes.dtype if torch.is_floating_point(bounding_boxes) else torch.float32
    device = bounding_boxes.device
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614

    # perspective_coeffs are computed as endpoint -> start point
    # We have to invert perspective_coeffs for bboxes:
    # (x, y) - end point and (x_out, y_out) - start point
    #   x_out = (coeffs[0] * x + coeffs[1] * y + coeffs[2]) / (coeffs[6] * x + coeffs[7] * y + 1)
    #   y_out = (coeffs[3] * x + coeffs[4] * y + coeffs[5]) / (coeffs[6] * x + coeffs[7] * y + 1)
    # and we would like to get:
    # x = (inv_coeffs[0] * x_out + inv_coeffs[1] * y_out + inv_coeffs[2])
    #       / (inv_coeffs[6] * x_out + inv_coeffs[7] * y_out + 1)
    # y = (inv_coeffs[3] * x_out + inv_coeffs[4] * y_out + inv_coeffs[5])
    #       / (inv_coeffs[6] * x_out + inv_coeffs[7] * y_out + 1)
    # and compute inv_coeffs in terms of coeffs

    denom = perspective_coeffs[0] * perspective_coeffs[4] - perspective_coeffs[1] * perspective_coeffs[3]
    if denom == 0:
        raise RuntimeError(
            f"Provided perspective_coeffs {perspective_coeffs} can not be inverted to transform bounding boxes. "
            f"Denominator is zero, denom={denom}"
        )

    inv_coeffs = [
        (perspective_coeffs[4] - perspective_coeffs[5] * perspective_coeffs[7]) / denom,
        (-perspective_coeffs[1] + perspective_coeffs[2] * perspective_coeffs[7]) / denom,
        (perspective_coeffs[1] * perspective_coeffs[5] - perspective_coeffs[2] * perspective_coeffs[4]) / denom,
        (-perspective_coeffs[3] + perspective_coeffs[5] * perspective_coeffs[6]) / denom,
        (perspective_coeffs[0] - perspective_coeffs[2] * perspective_coeffs[6]) / denom,
        (-perspective_coeffs[0] * perspective_coeffs[5] + perspective_coeffs[2] * perspective_coeffs[3]) / denom,
        (-perspective_coeffs[4] * perspective_coeffs[6] + perspective_coeffs[3] * perspective_coeffs[7]) / denom,
        (-perspective_coeffs[0] * perspective_coeffs[7] + perspective_coeffs[1] * perspective_coeffs[6]) / denom,
    ]

1615
1616
    theta1 = torch.tensor(
        [[inv_coeffs[0], inv_coeffs[1], inv_coeffs[2]], [inv_coeffs[3], inv_coeffs[4], inv_coeffs[5]]],
1617
1618
1619
1620
        dtype=dtype,
        device=device,
    )

1621
1622
1623
1624
    theta2 = torch.tensor(
        [[inv_coeffs[6], inv_coeffs[7], 1.0], [inv_coeffs[6], inv_coeffs[7], 1.0]], dtype=dtype, device=device
    )

1625
1626
1627
1628
    # 1) Let's transform bboxes into a tensor of 4 points (top-left, top-right, bottom-left, bottom-right corners).
    # Tensor of points has shape (N * 4, 3), where N is the number of bboxes
    # Single point structure is similar to
    # [(xmin, ymin, 1), (xmax, ymin, 1), (xmax, ymax, 1), (xmin, ymax, 1)]
1629
    points = bounding_boxes[:, [[0, 1], [2, 1], [2, 3], [0, 3]]].reshape(-1, 2)
1630
1631
1632
1633
1634
    points = torch.cat([points, torch.ones(points.shape[0], 1, device=points.device)], dim=-1)
    # 2) Now let's transform the points using perspective matrices
    #   x_out = (coeffs[0] * x + coeffs[1] * y + coeffs[2]) / (coeffs[6] * x + coeffs[7] * y + 1)
    #   y_out = (coeffs[3] * x + coeffs[4] * y + coeffs[5]) / (coeffs[6] * x + coeffs[7] * y + 1)

1635
1636
    numer_points = torch.matmul(points, theta1.T)
    denom_points = torch.matmul(points, theta2.T)
1637
    transformed_points = numer_points.div_(denom_points)
1638
1639
1640

    # 3) Reshape transformed points to [N boxes, 4 points, x/y coords]
    # and compute bounding box from 4 transformed points:
1641
    transformed_points = transformed_points.reshape(-1, 4, 2)
1642
1643
    out_bbox_mins, out_bbox_maxs = torch.aminmax(transformed_points, dim=1)

1644
1645
    out_bboxes = clamp_bounding_boxes(
        torch.cat([out_bbox_mins, out_bbox_maxs], dim=1).to(bounding_boxes.dtype),
1646
        format=datapoints.BoundingBoxFormat.XYXY,
Philip Meier's avatar
Philip Meier committed
1647
        canvas_size=canvas_size,
1648
    )
1649
1650
1651

    # out_bboxes should be of shape [N boxes, 4]

1652
    return convert_format_bounding_boxes(
1653
        out_bboxes, old_format=datapoints.BoundingBoxFormat.XYXY, new_format=format, inplace=True
1654
    ).reshape(original_shape)
1655
1656


1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
@_register_kernel_internal(perspective, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _perspective_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes,
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
    coefficients: Optional[List[float]] = None,
    **kwargs,
) -> datapoints.BoundingBoxes:
    output = perspective_bounding_boxes(
        inpt.as_subclass(torch.Tensor),
        format=inpt.format,
        canvas_size=inpt.canvas_size,
        startpoints=startpoints,
        endpoints=endpoints,
        coefficients=coefficients,
    )
1673
    return datapoints.wrap(output, like=inpt)
1674
1675


1676
1677
def perspective_mask(
    mask: torch.Tensor,
1678
1679
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
1680
    fill: _FillTypeJIT = None,
1681
    coefficients: Optional[List[float]] = None,
1682
) -> torch.Tensor:
1683
1684
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
1685
1686
1687
1688
1689
        needs_squeeze = True
    else:
        needs_squeeze = False

    output = perspective_image_tensor(
1690
        mask, startpoints, endpoints, interpolation=InterpolationMode.NEAREST, fill=fill, coefficients=coefficients
1691
    )
1692

1693
1694
1695
1696
1697
    if needs_squeeze:
        output = output.squeeze(0)

    return output

1698

1699
1700
1701
1702
1703
@_register_kernel_internal(perspective, datapoints.Mask, datapoint_wrapper=False)
def _perspective_mask_dispatch(
    inpt: datapoints.Mask,
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
1704
    fill: _FillTypeJIT = None,
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
    coefficients: Optional[List[float]] = None,
    **kwargs,
) -> datapoints.Mask:
    output = perspective_mask(
        inpt.as_subclass(torch.Tensor),
        startpoints=startpoints,
        endpoints=endpoints,
        fill=fill,
        coefficients=coefficients,
    )
1715
    return datapoints.wrap(output, like=inpt)
1716
1717
1718


@_register_kernel_internal(perspective, datapoints.Video)
1719
1720
def perspective_video(
    video: torch.Tensor,
1721
1722
    startpoints: Optional[List[List[int]]],
    endpoints: Optional[List[List[int]]],
1723
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1724
    fill: _FillTypeJIT = None,
1725
    coefficients: Optional[List[float]] = None,
1726
) -> torch.Tensor:
1727
1728
1729
    return perspective_image_tensor(
        video, startpoints, endpoints, interpolation=interpolation, fill=fill, coefficients=coefficients
    )
1730
1731


1732
def elastic(
1733
    inpt: torch.Tensor,
1734
    displacement: torch.Tensor,
1735
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1736
1737
    fill: _FillTypeJIT = None,
) -> torch.Tensor:
1738
1739
1740
1741
1742
1743
1744
    if torch.jit.is_scripting():
        return elastic_image_tensor(inpt, displacement=displacement, interpolation=interpolation, fill=fill)

    _log_api_usage_once(elastic)

    kernel = _get_kernel(elastic, type(inpt))
    return kernel(inpt, displacement=displacement, interpolation=interpolation, fill=fill)
1745
1746


1747
1748
1749
elastic_transform = elastic


1750
@_register_kernel_internal(elastic, torch.Tensor)
1751
@_register_kernel_internal(elastic, datapoints.Image)
1752
def elastic_image_tensor(
1753
    image: torch.Tensor,
1754
    displacement: torch.Tensor,
1755
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1756
    fill: _FillTypeJIT = None,
1757
) -> torch.Tensor:
Philip Meier's avatar
Philip Meier committed
1758
1759
1760
    if not isinstance(displacement, torch.Tensor):
        raise TypeError("Argument displacement should be a Tensor")

1761
1762
    interpolation = _check_interpolation(interpolation)

1763
1764
1765
1766
    if image.numel() == 0:
        return image

    shape = image.shape
1767
    ndim = image.ndim
1768

1769
    device = image.device
1770
    dtype = image.dtype if torch.is_floating_point(image) else torch.float32
1771
1772
1773
1774
1775
1776
1777

    # Patch: elastic transform should support (cpu,f16) input
    is_cpu_half = device.type == "cpu" and dtype == torch.float16
    if is_cpu_half:
        image = image.to(torch.float32)
        dtype = torch.float32

1778
1779
1780
    # We are aware that if input image dtype is uint8 and displacement is float64 then
    # displacement will be casted to float32 and all computations will be done with float32
    # We can fix this later if needed
1781

1782
1783
1784
1785
    expected_shape = (1,) + shape[-2:] + (2,)
    if expected_shape != displacement.shape:
        raise ValueError(f"Argument displacement shape should be {expected_shape}, but given {displacement.shape}")

1786
    if ndim > 4:
1787
        image = image.reshape((-1,) + shape[-3:])
1788
        needs_unsquash = True
1789
1790
1791
    elif ndim == 3:
        image = image.unsqueeze(0)
        needs_unsquash = True
1792
1793
1794
    else:
        needs_unsquash = False

1795
1796
    if displacement.dtype != dtype or displacement.device != device:
        displacement = displacement.to(dtype=dtype, device=device)
1797

1798
1799
1800
    image_height, image_width = shape[-2:]
    grid = _create_identity_grid((image_height, image_width), device=device, dtype=dtype).add_(displacement)
    output = _apply_grid_transform(image, grid, interpolation.value, fill=fill)
1801
1802

    if needs_unsquash:
1803
        output = output.reshape(shape)
1804

1805
1806
1807
    if is_cpu_half:
        output = output.to(torch.float16)

1808
    return output
1809
1810


1811
@_register_kernel_internal(elastic, PIL.Image.Image)
1812
def elastic_image_pil(
1813
    image: PIL.Image.Image,
1814
    displacement: torch.Tensor,
1815
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1816
    fill: _FillTypeJIT = None,
1817
) -> PIL.Image.Image:
1818
    t_img = pil_to_tensor(image)
1819
    output = elastic_image_tensor(t_img, displacement, interpolation=interpolation, fill=fill)
1820
    return to_pil_image(output, mode=image.mode)
1821
1822


1823
def _create_identity_grid(size: Tuple[int, int], device: torch.device, dtype: torch.dtype) -> torch.Tensor:
1824
    sy, sx = size
1825
1826
    base_grid = torch.empty(1, sy, sx, 2, device=device, dtype=dtype)
    x_grid = torch.linspace((-sx + 1) / sx, (sx - 1) / sx, sx, device=device, dtype=dtype)
1827
1828
    base_grid[..., 0].copy_(x_grid)

1829
    y_grid = torch.linspace((-sy + 1) / sy, (sy - 1) / sy, sy, device=device, dtype=dtype).unsqueeze_(-1)
1830
1831
1832
1833
1834
    base_grid[..., 1].copy_(y_grid)

    return base_grid


1835
1836
def elastic_bounding_boxes(
    bounding_boxes: torch.Tensor,
1837
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
1838
    canvas_size: Tuple[int, int],
1839
1840
    displacement: torch.Tensor,
) -> torch.Tensor:
Philip Meier's avatar
Philip Meier committed
1841
1842
1843
1844
1845
1846
    expected_shape = (1, canvas_size[0], canvas_size[1], 2)
    if not isinstance(displacement, torch.Tensor):
        raise TypeError("Argument displacement should be a Tensor")
    elif displacement.shape != expected_shape:
        raise ValueError(f"Argument displacement shape should be {expected_shape}, but given {displacement.shape}")

1847
1848
    if bounding_boxes.numel() == 0:
        return bounding_boxes
1849

1850
    # TODO: add in docstring about approximation we are doing for grid inversion
1851
1852
    device = bounding_boxes.device
    dtype = bounding_boxes.dtype if torch.is_floating_point(bounding_boxes) else torch.float32
1853
1854
1855

    if displacement.dtype != dtype or displacement.device != device:
        displacement = displacement.to(dtype=dtype, device=device)
1856

1857
1858
1859
1860
    original_shape = bounding_boxes.shape
    # TODO: first cast to float if bbox is int64 before convert_format_bounding_boxes
    bounding_boxes = (
        convert_format_bounding_boxes(bounding_boxes, old_format=format, new_format=datapoints.BoundingBoxFormat.XYXY)
1861
    ).reshape(-1, 4)
1862

Philip Meier's avatar
Philip Meier committed
1863
    id_grid = _create_identity_grid(canvas_size, device=device, dtype=dtype)
1864
1865
    # We construct an approximation of inverse grid as inv_grid = id_grid - displacement
    # This is not an exact inverse of the grid
1866
    inv_grid = id_grid.sub_(displacement)
1867
1868

    # Get points from bboxes
1869
    points = bounding_boxes[:, [[0, 1], [2, 1], [2, 3], [0, 3]]].reshape(-1, 2)
1870
1871
1872
1873
1874
    if points.is_floating_point():
        points = points.ceil_()
    index_xy = points.to(dtype=torch.long)
    index_x, index_y = index_xy[:, 0], index_xy[:, 1]

1875
    # Transform points:
Philip Meier's avatar
Philip Meier committed
1876
    t_size = torch.tensor(canvas_size[::-1], device=displacement.device, dtype=displacement.dtype)
1877
    transformed_points = inv_grid[0, index_y, index_x, :].add_(1).mul_(0.5 * t_size).sub_(0.5)
1878

1879
    transformed_points = transformed_points.reshape(-1, 4, 2)
1880
    out_bbox_mins, out_bbox_maxs = torch.aminmax(transformed_points, dim=1)
1881
1882
    out_bboxes = clamp_bounding_boxes(
        torch.cat([out_bbox_mins, out_bbox_maxs], dim=1).to(bounding_boxes.dtype),
1883
        format=datapoints.BoundingBoxFormat.XYXY,
Philip Meier's avatar
Philip Meier committed
1884
        canvas_size=canvas_size,
1885
    )
1886

1887
    return convert_format_bounding_boxes(
1888
        out_bboxes, old_format=datapoints.BoundingBoxFormat.XYXY, new_format=format, inplace=True
1889
    ).reshape(original_shape)
1890
1891


1892
1893
1894
1895
1896
1897
1898
@_register_kernel_internal(elastic, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _elastic_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, displacement: torch.Tensor, **kwargs
) -> datapoints.BoundingBoxes:
    output = elastic_bounding_boxes(
        inpt.as_subclass(torch.Tensor), format=inpt.format, canvas_size=inpt.canvas_size, displacement=displacement
    )
1899
    return datapoints.wrap(output, like=inpt)
1900
1901


1902
1903
1904
def elastic_mask(
    mask: torch.Tensor,
    displacement: torch.Tensor,
1905
    fill: _FillTypeJIT = None,
1906
) -> torch.Tensor:
1907
1908
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
1909
1910
1911
1912
        needs_squeeze = True
    else:
        needs_squeeze = False

1913
    output = elastic_image_tensor(mask, displacement=displacement, interpolation=InterpolationMode.NEAREST, fill=fill)
1914
1915
1916
1917
1918

    if needs_squeeze:
        output = output.squeeze(0)

    return output
1919
1920


1921
1922
@_register_kernel_internal(elastic, datapoints.Mask, datapoint_wrapper=False)
def _elastic_mask_dispatch(
1923
    inpt: datapoints.Mask, displacement: torch.Tensor, fill: _FillTypeJIT = None, **kwargs
1924
1925
) -> datapoints.Mask:
    output = elastic_mask(inpt.as_subclass(torch.Tensor), displacement=displacement, fill=fill)
1926
    return datapoints.wrap(output, like=inpt)
1927
1928
1929


@_register_kernel_internal(elastic, datapoints.Video)
1930
1931
1932
def elastic_video(
    video: torch.Tensor,
    displacement: torch.Tensor,
1933
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
1934
    fill: _FillTypeJIT = None,
1935
) -> torch.Tensor:
1936
    return elastic_image_tensor(video, displacement, interpolation=interpolation, fill=fill)
1937
1938


1939
def center_crop(inpt: torch.Tensor, output_size: List[int]) -> torch.Tensor:
1940
1941
1942
1943
1944
1945
1946
    if torch.jit.is_scripting():
        return center_crop_image_tensor(inpt, output_size=output_size)

    _log_api_usage_once(center_crop)

    kernel = _get_kernel(center_crop, type(inpt))
    return kernel(inpt, output_size=output_size)
1947
1948


1949
1950
def _center_crop_parse_output_size(output_size: List[int]) -> List[int]:
    if isinstance(output_size, numbers.Number):
1951
1952
        s = int(output_size)
        return [s, s]
1953
    elif isinstance(output_size, (tuple, list)) and len(output_size) == 1:
1954
        return [output_size[0], output_size[0]]
1955
1956
    else:
        return list(output_size)
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975


def _center_crop_compute_padding(crop_height: int, crop_width: int, image_height: int, image_width: int) -> List[int]:
    return [
        (crop_width - image_width) // 2 if crop_width > image_width else 0,
        (crop_height - image_height) // 2 if crop_height > image_height else 0,
        (crop_width - image_width + 1) // 2 if crop_width > image_width else 0,
        (crop_height - image_height + 1) // 2 if crop_height > image_height else 0,
    ]


def _center_crop_compute_crop_anchor(
    crop_height: int, crop_width: int, image_height: int, image_width: int
) -> Tuple[int, int]:
    crop_top = int(round((image_height - crop_height) / 2.0))
    crop_left = int(round((image_width - crop_width) / 2.0))
    return crop_top, crop_left


1976
@_register_kernel_internal(center_crop, torch.Tensor)
1977
@_register_kernel_internal(center_crop, datapoints.Image)
1978
def center_crop_image_tensor(image: torch.Tensor, output_size: List[int]) -> torch.Tensor:
1979
    crop_height, crop_width = _center_crop_parse_output_size(output_size)
1980
1981
1982
1983
    shape = image.shape
    if image.numel() == 0:
        return image.reshape(shape[:-2] + (crop_height, crop_width))
    image_height, image_width = shape[-2:]
1984
1985
1986

    if crop_height > image_height or crop_width > image_width:
        padding_ltrb = _center_crop_compute_padding(crop_height, crop_width, image_height, image_width)
1987
        image = torch_pad(image, _parse_pad_padding(padding_ltrb), value=0.0)
1988

1989
        image_height, image_width = image.shape[-2:]
1990
        if crop_width == image_width and crop_height == image_height:
1991
            return image
1992
1993

    crop_top, crop_left = _center_crop_compute_crop_anchor(crop_height, crop_width, image_height, image_width)
1994
    return image[..., crop_top : (crop_top + crop_height), crop_left : (crop_left + crop_width)]
1995
1996


1997
@_register_kernel_internal(center_crop, PIL.Image.Image)
1998
def center_crop_image_pil(image: PIL.Image.Image, output_size: List[int]) -> PIL.Image.Image:
1999
    crop_height, crop_width = _center_crop_parse_output_size(output_size)
Philip Meier's avatar
Philip Meier committed
2000
    image_height, image_width = get_size_image_pil(image)
2001
2002
2003

    if crop_height > image_height or crop_width > image_width:
        padding_ltrb = _center_crop_compute_padding(crop_height, crop_width, image_height, image_width)
2004
        image = pad_image_pil(image, padding_ltrb, fill=0)
2005

Philip Meier's avatar
Philip Meier committed
2006
        image_height, image_width = get_size_image_pil(image)
2007
        if crop_width == image_width and crop_height == image_height:
2008
            return image
2009
2010

    crop_top, crop_left = _center_crop_compute_crop_anchor(crop_height, crop_width, image_height, image_width)
2011
    return crop_image_pil(image, crop_top, crop_left, crop_height, crop_width)
2012
2013


2014
2015
def center_crop_bounding_boxes(
    bounding_boxes: torch.Tensor,
2016
    format: datapoints.BoundingBoxFormat,
Philip Meier's avatar
Philip Meier committed
2017
    canvas_size: Tuple[int, int],
2018
    output_size: List[int],
2019
) -> Tuple[torch.Tensor, Tuple[int, int]]:
2020
    crop_height, crop_width = _center_crop_parse_output_size(output_size)
Philip Meier's avatar
Philip Meier committed
2021
    crop_top, crop_left = _center_crop_compute_crop_anchor(crop_height, crop_width, *canvas_size)
2022
2023
2024
    return crop_bounding_boxes(
        bounding_boxes, format, top=crop_top, left=crop_left, height=crop_height, width=crop_width
    )
2025
2026


2027
2028
2029
2030
2031
2032
2033
@_register_kernel_internal(center_crop, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _center_crop_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, output_size: List[int]
) -> datapoints.BoundingBoxes:
    output, canvas_size = center_crop_bounding_boxes(
        inpt.as_subclass(torch.Tensor), format=inpt.format, canvas_size=inpt.canvas_size, output_size=output_size
    )
2034
    return datapoints.wrap(output, like=inpt, canvas_size=canvas_size)
2035
2036
2037


@_register_kernel_internal(center_crop, datapoints.Mask)
2038
2039
2040
def center_crop_mask(mask: torch.Tensor, output_size: List[int]) -> torch.Tensor:
    if mask.ndim < 3:
        mask = mask.unsqueeze(0)
2041
2042
2043
2044
        needs_squeeze = True
    else:
        needs_squeeze = False

2045
    output = center_crop_image_tensor(image=mask, output_size=output_size)
2046
2047
2048
2049
2050

    if needs_squeeze:
        output = output.squeeze(0)

    return output
2051
2052


2053
@_register_kernel_internal(center_crop, datapoints.Video)
2054
2055
2056
2057
def center_crop_video(video: torch.Tensor, output_size: List[int]) -> torch.Tensor:
    return center_crop_image_tensor(video, output_size)


2058
def resized_crop(
2059
    inpt: torch.Tensor,
2060
2061
2062
2063
2064
2065
2066
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
    antialias: Optional[Union[str, bool]] = "warn",
2067
) -> torch.Tensor:
2068
    if torch.jit.is_scripting():
2069
        return resized_crop_image_tensor(
2070
2071
2072
2073
2074
2075
2076
2077
            inpt,
            top=top,
            left=left,
            height=height,
            width=width,
            size=size,
            interpolation=interpolation,
            antialias=antialias,
2078
        )
2079

2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
    _log_api_usage_once(resized_crop)

    kernel = _get_kernel(resized_crop, type(inpt))
    return kernel(
        inpt,
        top=top,
        left=left,
        height=height,
        width=width,
        size=size,
        interpolation=interpolation,
        antialias=antialias,
    )
2093

2094
2095

@_register_kernel_internal(resized_crop, torch.Tensor)
2096
@_register_kernel_internal(resized_crop, datapoints.Image)
2097
def resized_crop_image_tensor(
2098
    image: torch.Tensor,
2099
2100
2101
2102
2103
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
2104
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
2105
    antialias: Optional[Union[str, bool]] = "warn",
2106
) -> torch.Tensor:
2107
2108
    image = crop_image_tensor(image, top, left, height, width)
    return resize_image_tensor(image, size, interpolation=interpolation, antialias=antialias)
2109
2110
2111


def resized_crop_image_pil(
2112
    image: PIL.Image.Image,
2113
2114
2115
2116
2117
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
2118
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
2119
) -> PIL.Image.Image:
2120
2121
    image = crop_image_pil(image, top, left, height, width)
    return resize_image_pil(image, size, interpolation=interpolation)
2122
2123


2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
@_register_kernel_internal(resized_crop, PIL.Image.Image)
def resized_crop_image_pil_dispatch(
    image: PIL.Image.Image,
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
    antialias: Optional[Union[str, bool]] = "warn",
) -> PIL.Image.Image:
    if antialias is False:
        warnings.warn("Anti-alias option is always applied for PIL Image input. Argument antialias is ignored.")
    return resized_crop_image_pil(
        image,
        top=top,
        left=left,
        height=height,
        width=width,
        size=size,
        interpolation=interpolation,
    )


2148
2149
def resized_crop_bounding_boxes(
    bounding_boxes: torch.Tensor,
2150
    format: datapoints.BoundingBoxFormat,
2151
2152
2153
2154
2155
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
2156
) -> Tuple[torch.Tensor, Tuple[int, int]]:
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
    bounding_boxes, canvas_size = crop_bounding_boxes(bounding_boxes, format, top, left, height, width)
    return resize_bounding_boxes(bounding_boxes, canvas_size=canvas_size, size=size)


@_register_kernel_internal(resized_crop, datapoints.BoundingBoxes, datapoint_wrapper=False)
def _resized_crop_bounding_boxes_dispatch(
    inpt: datapoints.BoundingBoxes, top: int, left: int, height: int, width: int, size: List[int], **kwargs
) -> datapoints.BoundingBoxes:
    output, canvas_size = resized_crop_bounding_boxes(
        inpt.as_subclass(torch.Tensor), format=inpt.format, top=top, left=left, height=height, width=width, size=size
    )
2168
    return datapoints.wrap(output, like=inpt, canvas_size=canvas_size)
2169
2170


2171
def resized_crop_mask(
2172
2173
2174
2175
2176
2177
2178
    mask: torch.Tensor,
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
) -> torch.Tensor:
2179
2180
    mask = crop_mask(mask, top, left, height, width)
    return resize_mask(mask, size)
2181
2182


2183
2184
2185
2186
2187
2188
2189
@_register_kernel_internal(resized_crop, datapoints.Mask, datapoint_wrapper=False)
def _resized_crop_mask_dispatch(
    inpt: datapoints.Mask, top: int, left: int, height: int, width: int, size: List[int], **kwargs
) -> datapoints.Mask:
    output = resized_crop_mask(
        inpt.as_subclass(torch.Tensor), top=top, left=left, height=height, width=width, size=size
    )
2190
    return datapoints.wrap(output, like=inpt)
2191
2192
2193


@_register_kernel_internal(resized_crop, datapoints.Video)
2194
2195
2196
2197
2198
2199
2200
def resized_crop_video(
    video: torch.Tensor,
    top: int,
    left: int,
    height: int,
    width: int,
    size: List[int],
2201
    interpolation: Union[InterpolationMode, int] = InterpolationMode.BILINEAR,
2202
    antialias: Optional[Union[str, bool]] = "warn",
2203
2204
2205
2206
2207
2208
) -> torch.Tensor:
    return resized_crop_image_tensor(
        video, top, left, height, width, antialias=antialias, size=size, interpolation=interpolation
    )


2209
def five_crop(
2210
2211
    inpt: torch.Tensor, size: List[int]
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
2212
2213
2214
2215
2216
2217
2218
    if torch.jit.is_scripting():
        return five_crop_image_tensor(inpt, size=size)

    _log_api_usage_once(five_crop)

    kernel = _get_kernel(five_crop, type(inpt))
    return kernel(inpt, size=size)
2219
2220


2221
2222
def _parse_five_crop_size(size: List[int]) -> List[int]:
    if isinstance(size, numbers.Number):
2223
2224
        s = int(size)
        size = [s, s]
2225
    elif isinstance(size, (tuple, list)) and len(size) == 1:
2226
2227
        s = size[0]
        size = [s, s]
2228
2229
2230
2231
2232
2233
2234

    if len(size) != 2:
        raise ValueError("Please provide only two dimensions (h, w) for size.")

    return size


2235
2236
@_register_five_ten_crop_kernel_internal(five_crop, torch.Tensor)
@_register_five_ten_crop_kernel_internal(five_crop, datapoints.Image)
2237
def five_crop_image_tensor(
2238
    image: torch.Tensor, size: List[int]
2239
2240
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    crop_height, crop_width = _parse_five_crop_size(size)
2241
    image_height, image_width = image.shape[-2:]
2242
2243

    if crop_width > image_width or crop_height > image_height:
2244
        raise ValueError(f"Requested crop size {size} is bigger than input size {(image_height, image_width)}")
2245

2246
2247
2248
2249
2250
    tl = crop_image_tensor(image, 0, 0, crop_height, crop_width)
    tr = crop_image_tensor(image, 0, image_width - crop_width, crop_height, crop_width)
    bl = crop_image_tensor(image, image_height - crop_height, 0, crop_height, crop_width)
    br = crop_image_tensor(image, image_height - crop_height, image_width - crop_width, crop_height, crop_width)
    center = center_crop_image_tensor(image, [crop_height, crop_width])
2251
2252
2253
2254

    return tl, tr, bl, br, center


2255
@_register_five_ten_crop_kernel_internal(five_crop, PIL.Image.Image)
2256
def five_crop_image_pil(
2257
    image: PIL.Image.Image, size: List[int]
2258
2259
) -> Tuple[PIL.Image.Image, PIL.Image.Image, PIL.Image.Image, PIL.Image.Image, PIL.Image.Image]:
    crop_height, crop_width = _parse_five_crop_size(size)
Philip Meier's avatar
Philip Meier committed
2260
    image_height, image_width = get_size_image_pil(image)
2261
2262

    if crop_width > image_width or crop_height > image_height:
2263
        raise ValueError(f"Requested crop size {size} is bigger than input size {(image_height, image_width)}")
2264

2265
2266
2267
2268
2269
    tl = crop_image_pil(image, 0, 0, crop_height, crop_width)
    tr = crop_image_pil(image, 0, image_width - crop_width, crop_height, crop_width)
    bl = crop_image_pil(image, image_height - crop_height, 0, crop_height, crop_width)
    br = crop_image_pil(image, image_height - crop_height, image_width - crop_width, crop_height, crop_width)
    center = center_crop_image_pil(image, [crop_height, crop_width])
2270
2271
2272
2273

    return tl, tr, bl, br, center


2274
@_register_five_ten_crop_kernel_internal(five_crop, datapoints.Video)
2275
2276
2277
2278
2279
2280
def five_crop_video(
    video: torch.Tensor, size: List[int]
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    return five_crop_image_tensor(video, size)


2281
def ten_crop(
2282
    inpt: torch.Tensor, size: List[int], vertical_flip: bool = False
2283
) -> Tuple[
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
2294
]:
2295
2296
2297
2298
2299
2300
2301
    if torch.jit.is_scripting():
        return ten_crop_image_tensor(inpt, size=size, vertical_flip=vertical_flip)

    _log_api_usage_once(ten_crop)

    kernel = _get_kernel(ten_crop, type(inpt))
    return kernel(inpt, size=size, vertical_flip=vertical_flip)
2302
2303


2304
2305
@_register_five_ten_crop_kernel_internal(ten_crop, torch.Tensor)
@_register_five_ten_crop_kernel_internal(ten_crop, datapoints.Image)
Philip Meier's avatar
Philip Meier committed
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
def ten_crop_image_tensor(
    image: torch.Tensor, size: List[int], vertical_flip: bool = False
) -> Tuple[
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
]:
    non_flipped = five_crop_image_tensor(image, size)
2321
2322

    if vertical_flip:
2323
        image = vertical_flip_image_tensor(image)
2324
    else:
2325
        image = horizontal_flip_image_tensor(image)
2326

Philip Meier's avatar
Philip Meier committed
2327
    flipped = five_crop_image_tensor(image, size)
2328

Philip Meier's avatar
Philip Meier committed
2329
    return non_flipped + flipped
2330
2331


2332
@_register_five_ten_crop_kernel_internal(ten_crop, PIL.Image.Image)
Philip Meier's avatar
Philip Meier committed
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
def ten_crop_image_pil(
    image: PIL.Image.Image, size: List[int], vertical_flip: bool = False
) -> Tuple[
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
    PIL.Image.Image,
]:
    non_flipped = five_crop_image_pil(image, size)
2348
2349

    if vertical_flip:
2350
        image = vertical_flip_image_pil(image)
2351
    else:
2352
        image = horizontal_flip_image_pil(image)
2353

Philip Meier's avatar
Philip Meier committed
2354
2355
2356
2357
2358
    flipped = five_crop_image_pil(image, size)

    return non_flipped + flipped


2359
@_register_five_ten_crop_kernel_internal(ten_crop, datapoints.Video)
Philip Meier's avatar
Philip Meier committed
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
def ten_crop_video(
    video: torch.Tensor, size: List[int], vertical_flip: bool = False
) -> Tuple[
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
    torch.Tensor,
]:
2374
    return ten_crop_image_tensor(video, size, vertical_flip=vertical_flip)