File size: 19,220 Bytes
8f72b1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
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
552
553
554
555
556
557
558
"""#TODO: dont convert to numpy and back to torch."""

from collections.abc import Iterable, Sequence
from itertools import chain
from typing import Any

import kornia.augmentation as K
import numpy as np
import torch
from kornia.augmentation import random_generator as rg
from kornia.augmentation.utils import _range_bound
from kornia.constants import DataKey, Resample
from typing import Optional, Tuple, Sequence, Dict, Union

def default_augmenter(coords: np.ndarray):
    # TODO parametrize magnitude of different augmentations
    ndim = coords.shape[1]

    assert coords.ndim == 2 and ndim in (2, 3)

    # first remove offset
    center = coords.mean(axis=0, keepdims=True)

    coords = coords - center

    # apply random flip
    coords *= 2 * np.random.randint(0, 2, (1, ndim)) - 1

    # apply rotation along the last two dimensions
    phi = np.random.uniform(0, 2 * np.pi)
    coords = _rotate(coords, phi, center=None)

    if ndim == 3:
        # rotate along the first two dimensions too
        phi2, phi3 = np.random.uniform(0, 2 * np.pi, 2)
        coords = _rotate(coords, phi2, rot_axis=(0, 1), center=None)
        coords = _rotate(coords, phi3, rot_axis=(0, 2), center=None)

    coords += center

    # translation
    trans = 128 * np.random.uniform(-1, 1, (1, ndim))
    coords += trans

    # elastic
    coords += 1.5 * np.random.normal(0, 1, coords.shape)

    return coords


def _rotate(
    coords: np.ndarray, phi: float, rot_axis=(-2, -1), center: Optional[Tuple] = None
):
    """Rotation along the last two dimensions of coords[..,:-2:]."""
    ndim = coords.shape[1]
    assert coords.ndim == 2 and ndim in (2, 3)

    if center is None:
        center = (0,) * ndim

    assert len(center) == ndim

    center = np.asarray(center)
    co, si = np.cos(phi), np.sin(phi)
    Rot = np.eye(ndim)
    Rot[np.ix_(rot_axis, rot_axis)] = np.array(((co, -si), (si, co)))
    x = coords - center
    x = x @ Rot.T
    x += center
    return x


def _filter_points(
    points: np.ndarray, shape: tuple, origin: Optional[Tuple] = None
) -> np.ndarray:
    """Returns indices of points that are inside the shape extent and given origin."""
    ndim = points.shape[-1]
    if origin is None:
        origin = (0,) * ndim

    idx = tuple(
        np.logical_and(points[:, i] >= origin[i], points[:, i] < origin[i] + shape[i])
        for i in range(ndim)
    )
    idx = np.where(np.all(idx, axis=0))[0]
    return idx


class ConcatAffine(K.RandomAffine):
    """Concatenate multiple affine transformations without intermediates."""

    def __init__(self, affines: Sequence[K.RandomAffine]):
        super().__init__(degrees=0)
        self._affines = affines
        if not all([a.same_on_batch for a in affines]):
            raise ValueError("all affines must have same_on_batch=True")

    def merge_params(self, params: Sequence[Dict[str, torch.Tensor]]):
        """Merge params from affines."""
        out = params[0].copy()

        def _torchmax(x, dim):
            return torch.max(x, dim=dim).values

        ops = {
            "translations": torch.sum,
            "center": torch.mean,
            "scale": torch.prod,
            "shear_x": torch.sum,
            "shear_y": torch.sum,
            "angle": torch.sum,
            "batch_prob": _torchmax,
        }
        for k, v in params[0].items():
            ps = [p[k] for p in params if len(p[k]) > 0]
            if len(ps) > 0 and k in ops:
                v_new = torch.stack(ps, dim=0).float()
                v_new = ops[k](v_new, dim=0)
                v_new = v_new.to(v.dtype)
            else:
                v_new = v
            out[k] = v_new

        return out

    def forward_parameters(
        self, batch_shape: Tuple[int, ...]
    ) -> Dict[str, torch.Tensor]:
        params = tuple(a.forward_parameters(batch_shape) for a in self._affines)
        # print(params)
        return self.merge_params(params)


# custom augmentations
class RandomIntensityScaleShift(K.IntensityAugmentationBase2D):
    r"""Apply a random scale and shift to the image intensity.

    Args:
        p: probability of applying the transformation.
        scale:  the scale factor to apply
        shift: the offset to apply
        clip_output: if true clip output
        same_on_batch: apply the same transformation across the batch.
        keepdim: whether to keep the output shape the same as input (True) or broadcast it
                 to the batch form (False).
    Shape:
        - Input: :math:`(C, H, W)` or :math:`(B, C, H, W)`, Optional: :math:`(B, 3, 3)`
        - Output: :math:`(B, C, H, W)`

    .. note::
        This function internally uses :func:`kornia.enhance.adjust_brightness`

    """

    def __init__(
        self,
        scale: Tuple[float, float] = (0.5, 2.0),
        shift: Tuple[float, float] = (-0.1, 0.1),
        clip_output: bool = True,
        same_on_batch: bool = False,
        p: float = 1.0,
        keepdim: bool = False,
    ) -> None:
        super().__init__(p=p, same_on_batch=same_on_batch, keepdim=keepdim)
        self.scale = _range_bound(
            scale, "scale", center=0, bounds=(-float("inf"), float("inf"))
        )
        self.shift = _range_bound(
            shift, "shift", center=0, bounds=(-float("inf"), float("inf"))
        )
        self._param_generator = rg.PlainUniformGenerator(
            (self.scale, "scale_factor", None, None),
            (self.shift, "shift_factor", None, None),
        )

        self.clip_output = clip_output

    def apply_transform(
        self,
        input: torch.Tensor,
        params: Dict[str, torch.Tensor],
        flags: Dict[str, Any],
        transform: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        scale_factor = params["scale_factor"].to(input)
        shift_factor = params["shift_factor"].to(input)
        scale_factor = scale_factor.view(len(scale_factor), 1, 1, 1)
        shift_factor = shift_factor.view(len(scale_factor), 1, 1, 1)
        img_adjust = input * scale_factor + shift_factor
        if self.clip_output:
            img_adjust = img_adjust.clamp(min=0.0, max=1.0)
        return img_adjust


class RandomTemporalAffine(K.RandomAffine):
    r"""Apply a random 2D affine transformation to a batch of images while
    varying the transformation across the time dimension from 0 to 1.

    Same args/kwargs as K.RandomAffine

    """

    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, same_on_batch=True, **kwargs)

    def forward_parameters(
        self, batch_shape: Tuple[int, ...]
    ) -> Dict[str, torch.Tensor]:
        params = super().forward_parameters(batch_shape)
        factor = torch.linspace(0, 1, batch_shape[0]).to(params["translations"])
        for key in ["translations", "center", "angle", "shear_x", "shear_y"]:
            v = params[key]
            if len(v) > 0:
                params[key] = v * factor.view(*((-1,) + (1,) * len(v.shape[1:])))

        for key in [
            "scale",
        ]:
            v = params[key]
            if len(v) > 0:
                params[key] = 1 + (v - 1) * factor.view(
                    *((-1,) + (1,) * len(v.shape[1:]))
                )
        return params

    # def compute_transformation(self, input: torch.Tensor,
    #                            params: Dict[str, torch.Tensor],
    #                            flags: Dict[str, Any]) -> torch.Tensor:
    #     factor = torch.linspace(0, 1, input.shape[0]).to(input)
    #     for key in ["translations", "center", "angle", "shear_x", "shear_y"]:
    #         v = params[key]
    #         params[key] = v * factor.view(*((-1,)+(1,)*len(v.shape[1:])))

    #     for key in ["scale", ]:
    #         v = params[key]
    #         params[key] = 1 + (v-1) * factor.view(*((-1,)+(1,)*len(v.shape[1:])))

    #     return super().compute_transformation(input, params, flags)


class BasicPipeline:
    """transforms img, mask, and points.

    Only supports 2D transformations for now (any 3D object will preserve its z coordinates/dimensions)
    """

    def __init__(self, augs: tuple, filter_points: bool = True):
        self.data_keys = ("input", "mask", "keypoints")
        self.pipeline = K.AugmentationSequential(
            *augs,
            # disable align_corners to not trigger lots of warnings from kornia
            extra_args={
                DataKey.MASK: {"resample": Resample.NEAREST, "align_corners": False}
            },
            data_keys=self.data_keys,
        )
        self.filter_points = filter_points

    def __call__(
        self,
        img: np.ndarray,
        mask: np.ndarray,
        points: np.ndarray,
        timepoints: np.ndarray,
    ):
        ndim = img.ndim - 1
        assert (
            ndim in (2, 3)
            and points.ndim == 2
            and points.shape[-1] == ndim
            and timepoints.ndim == 1
            and img.shape == mask.shape
        )

        x = torch.from_numpy(img).float()
        y = torch.from_numpy(mask.astype(np.int64)).float()

        # if 2D add dummy channel
        if ndim == 2:
            x = x.unsqueeze(1)
            y = y.unsqueeze(1)
            p = points[..., [1, 0]]
        # if 3D we use z as channel (i.e. fix augs across z)
        elif ndim == 3:
            p = points[..., [2, 1]]

        # flip as kornia expects xy and not yx
        p = torch.from_numpy(p).unsqueeze(0).float()
        # add batch by duplicating to make kornia happy
        p = p.expand(len(x), -1, -1)
        # create a mask to know which timepoint the points belong to
        ts = torch.from_numpy(timepoints).long()
        n_points = p.shape[1]
        if n_points > 0:
            x, y, p = self.pipeline(x, y, p)
        else:
            # dummy keypoints
            x, y = self.pipeline(x, y, torch.zeros((len(x), 1, 2)))[:2]

        # remove batch
        p = p[ts, torch.arange(n_points)]
        # flip back
        p = p[..., [1, 0]]

        # remove channel
        if ndim == 2:
            x = x.squeeze(1)
            y = y.squeeze(1)

        x = x.numpy()
        y = y.numpy().astype(np.uint16)
        # p = p.squeeze(0).numpy()
        p = p.numpy()
        # add back z coordinates
        if ndim == 3:
            p = np.concatenate([points[..., 0:1], p], axis=-1)
        ts = ts.numpy()
        # remove points outside of img/mask

        if self.filter_points:
            idx = _filter_points(p, shape=x.shape[-ndim:])

        else:
            idx = np.arange(len(p), dtype=int)

        p = p[idx]
        return (x, y, p), idx


class RandomCrop:
    def __init__(
        self,
        crop_size: Optional[Union[int, Tuple[int]]] = None,
        ndim: int = 2,
        ensure_inside_points: bool = False,
        use_padding: bool = True,
        padding_mode="constant",
    ) -> None:
        """crop_size: tuple of int
        can be tuple of length 1 (all dimensions)
                     of length ndim (y,x,...)
                     of length 2*ndim (y1,y2, x1,x2, ...).
        """
        if isinstance(crop_size, int):
            crop_size = (crop_size,) * 2 * ndim
        elif isinstance(crop_size, Iterable):
            pass
        else:
            raise ValueError(f"{crop_size} has to be int or tuple of int")

        if len(crop_size) == 1:
            crop_size = (crop_size[0],) * 2 * ndim
        elif len(crop_size) == ndim:
            crop_size = tuple(chain(*tuple((c, c) for c in crop_size)))
        elif len(crop_size) == 2 * ndim:
            pass
        else:
            raise ValueError(f"crop_size has to be of length 1, {ndim}, or {2 * ndim}")

        crop_size = np.array(crop_size)
        self._ndim = ndim
        self._crop_bounds = crop_size[::2], crop_size[1::2]
        self._use_padding = use_padding
        self._ensure_inside_points = ensure_inside_points
        self._rng = np.random.RandomState()
        self._padding_mode = padding_mode

    def crop_img(self, img: np.ndarray, corner: np.ndarray, crop_size: np.ndarray):
        if not img.ndim == self._ndim + 1:
            raise ValueError(
                f"img has to be 1 (time) + {self._ndim} spatial dimensions"
            )

        pad_left = np.maximum(0, -corner)
        pad_right = np.maximum(
            0, corner + crop_size - np.array(img.shape[-self._ndim :])
        )

        img = np.pad(
            img,
            ((0, 0), *tuple(np.stack((pad_left, pad_right)).T)),
            mode=self._padding_mode,
        )
        slices = (
            slice(None),
            *tuple(slice(c, c + s) for c, s in zip(corner + pad_left, crop_size)),
        )
        return img[slices]

    def crop_points(
        self, points: np.ndarray, corner: np.ndarray, crop_size: np.ndarray
    ):
        idx = _filter_points(points, shape=crop_size, origin=corner)
        return points[idx] - corner, idx

    def __call__(self, img: np.ndarray, mask: np.ndarray, points: np.ndarray):
        assert (
            img.ndim == self._ndim + 1
            and points.ndim == 2
            and points.shape[-1] == self._ndim
            and img.shape == mask.shape
        )

        points = points.astype(int)

        crop_size = self._rng.randint(self._crop_bounds[0], self._crop_bounds[1] + 1)
        # print(f'{crop_size=}')

        if self._ensure_inside_points:
            if len(points) == 0:
                print("No points given, cannot ensure inside points")
                return (img, mask, points), np.zeros((0,), int)

            # sample point and corner relative to it

            _idx = np.random.randint(len(points))
            corner = (
                points[_idx]
                - crop_size
                + 1
                + self._rng.randint(crop_size // 4, 3 * crop_size // 4)
            )
        else:
            corner = self._rng.randint(
                0, np.maximum(1, np.array(img.shape[-self._ndim :]) - crop_size)
            )

        if not self._use_padding:
            corner = np.maximum(0, corner)
            crop_size = np.minimum(
                crop_size, np.array(img.shape[-self._ndim :]) - corner
            )

        img = self.crop_img(img, corner, crop_size)
        mask = self.crop_img(mask, corner, crop_size)
        points, idx = self.crop_points(points, corner, crop_size)

        return (img, mask, points), idx


class AugmentationPipeline(BasicPipeline):
    """transforms img, mask, and points."""

    def __init__(self, p=0.5, filter_points=True, level=1):
        if level == 1:
            augs = [
                # Augmentations for all images in a window
                K.RandomHorizontalFlip(p=0.5, same_on_batch=True),
                K.RandomVerticalFlip(p=0.5, same_on_batch=True),
                K.RandomAffine(
                    degrees=180,
                    shear=(-10, 10, -10, 10),  # x_min, x_max, y_min, y_max
                    translate=(0.05, 0.05),
                    scale=(0.8, 1.2),  # x_min, x_max, y_min, y_max
                    p=p,
                    same_on_batch=True,
                ),
                K.RandomBrightness(
                    (0.5, 1.5), clip_output=False, p=p, same_on_batch=True
                ),
                K.RandomGaussianNoise(mean=0.0, std=0.03, p=p, same_on_batch=False),
            ]
        elif level == 2:
            # Crafted for DeepCell crop size 256
            augs = [
                # Augmentations for all images in a window
                K.RandomHorizontalFlip(p=0.5, same_on_batch=True),
                K.RandomVerticalFlip(p=0.5, same_on_batch=True),
                K.RandomAffine(
                    degrees=180,
                    shear=(-5, 5, -5, 5),  # x_min, x_max, y_min, y_max
                    translate=(0.03, 0.03),
                    scale=(0.8, 1.2),  # isotropic
                    p=p,
                    same_on_batch=True,
                ),
                # Anisotropic scaling
                K.RandomAffine(
                    degrees=0,
                    scale=(0.9, 1.1, 0.9, 1.1),  # x_min, x_max, y_min, y_max
                    p=p,
                    same_on_batch=True,
                ),
                # Independet augmentations for each image in window
                K.RandomAffine(
                    degrees=3,
                    shear=(-2, 2, -2, 2),  # x_min, x_max, y_min, y_max
                    translate=(0.04, 0.04),
                    p=p,
                    same_on_batch=False,
                ),
                # not implemented for points in kornia 0.7.0
                # K.RandomElasticTransform(alpha=50, sigma=5, p=p, same_on_batch=False),
                # Intensity-based augmentations
                K.RandomBrightness(
                    (0.5, 1.5), clip_output=False, p=p, same_on_batch=True
                ),
                K.RandomGaussianNoise(mean=0.0, std=0.03, p=p, same_on_batch=False),
            ]
        elif level == 3:
            # Crafted for DeepCell crop size 256
            augs = [
                # Augmentations for all images in a window
                K.RandomHorizontalFlip(p=0.5, same_on_batch=True),
                K.RandomVerticalFlip(p=0.5, same_on_batch=True),
                ConcatAffine([
                    K.RandomAffine(
                        degrees=180,
                        shear=(-5, 5, -5, 5),  # x_min, x_max, y_min, y_max
                        translate=(0.03, 0.03),
                        scale=(0.8, 1.2),  # isotropic
                        p=p,
                        same_on_batch=True,
                    ),
                    # Anisotropic scaling
                    K.RandomAffine(
                        degrees=0,
                        scale=(0.9, 1.1, 0.9, 1.1),  # x_min, x_max, y_min, y_max
                        p=p,
                        same_on_batch=True,
                    ),
                ]),
                RandomTemporalAffine(
                    degrees=10,
                    translate=(0.05, 0.05),
                    p=p,
                    # same_on_batch=True,
                ),
                # Independet augmentations for each image in window
                K.RandomAffine(
                    degrees=2,
                    shear=(-2, 2, -2, 2),  # x_min, x_max, y_min, y_max
                    translate=(0.01, 0.01),
                    p=0.5 * p,
                    same_on_batch=False,
                ),
                # Intensity-based augmentations
                RandomIntensityScaleShift(
                    (0.5, 2.0), (-0.1, 0.1), clip_output=False, p=p, same_on_batch=True
                ),
                K.RandomGaussianNoise(mean=0.0, std=0.03, p=p, same_on_batch=False),
            ]
        elif level == 4:
            # debug
            augs = [
                K.RandomAffine(
                    degrees=30,
                    shear=(-0, 0, -0, 0),  # x_min, x_max, y_min, y_max
                    translate=(0.0, 0.0),
                    p=1,
                    same_on_batch=True,
                ),
            ]
        else:
            raise ValueError(f"level {level} not supported")

        super().__init__(augs, filter_points)