test_functional_tensor.py 28.2 KB
Newer Older
1
import unittest
2
import colorsys
3
import math
4

vfdev's avatar
vfdev committed
5
import numpy as np
6
from PIL.Image import NEAREST, BILINEAR, BICUBIC
vfdev's avatar
vfdev committed
7
8
9
10
11
12

import torch
import torchvision.transforms as transforms
import torchvision.transforms.functional_tensor as F_t
import torchvision.transforms.functional_pil as F_pil
import torchvision.transforms.functional as F
13

14
from common_utils import TransformsTester
15

16

17
class Tester(TransformsTester):
vfdev's avatar
vfdev committed
18

19
    def _test_vflip(self, device):
20
        script_vflip = torch.jit.script(F_t.vflip)
21
        img_tensor = torch.randn(3, 16, 16, device=device)
22
        img_tensor_clone = img_tensor.clone()
23
24
25
26
        vflipped_img = F_t.vflip(img_tensor)
        vflipped_img_again = F_t.vflip(vflipped_img)
        self.assertEqual(vflipped_img.shape, img_tensor.shape)
        self.assertTrue(torch.equal(img_tensor, vflipped_img_again))
27
        self.assertTrue(torch.equal(img_tensor, img_tensor_clone))
28
29
30
        # scriptable function test
        vflipped_img_script = script_vflip(img_tensor)
        self.assertTrue(torch.equal(vflipped_img, vflipped_img_script))
31

32
33
34
35
36
37
38
39
    def test_vflip_cpu(self):
        self._test_vflip("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_vflip_cuda(self):
        self._test_vflip("cuda")

    def _test_hflip(self, device):
40
        script_hflip = torch.jit.script(F_t.hflip)
41
        img_tensor = torch.randn(3, 16, 16, device=device)
42
        img_tensor_clone = img_tensor.clone()
43
44
45
46
        hflipped_img = F_t.hflip(img_tensor)
        hflipped_img_again = F_t.hflip(hflipped_img)
        self.assertEqual(hflipped_img.shape, img_tensor.shape)
        self.assertTrue(torch.equal(img_tensor, hflipped_img_again))
47
        self.assertTrue(torch.equal(img_tensor, img_tensor_clone))
48
49
50
        # scriptable function test
        hflipped_img_script = script_hflip(img_tensor)
        self.assertTrue(torch.equal(hflipped_img, hflipped_img_script))
51

52
53
54
55
56
57
58
59
60
    def test_hflip_cpu(self):
        self._test_hflip("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_hflip_cuda(self):
        self._test_hflip("cuda")

    def _test_crop(self, device):
        script_crop = torch.jit.script(F.crop)
61

62
        img_tensor, pil_img = self._create_data(16, 18, device=device)
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78

        test_configs = [
            (1, 2, 4, 5),   # crop inside top-left corner
            (2, 12, 3, 4),  # crop inside top-right corner
            (8, 3, 5, 6),   # crop inside bottom-left corner
            (8, 11, 4, 3),  # crop inside bottom-right corner
        ]

        for top, left, height, width in test_configs:
            pil_img_cropped = F.crop(pil_img, top, left, height, width)

            img_tensor_cropped = F.crop(img_tensor, top, left, height, width)
            self.compareTensorToPIL(img_tensor_cropped, pil_img_cropped)

            img_tensor_cropped = script_crop(img_tensor, top, left, height, width)
            self.compareTensorToPIL(img_tensor_cropped, pil_img_cropped)
ekka's avatar
ekka committed
79

80
81
82
83
84
85
86
    def test_crop_cpu(self):
        self._test_crop("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_crop_cuda(self):
        self._test_crop("cuda")

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
    def test_hsv2rgb(self):
        shape = (3, 100, 150)
        for _ in range(20):
            img = torch.rand(*shape, dtype=torch.float)
            ft_img = F_t._hsv2rgb(img).permute(1, 2, 0).flatten(0, 1)

            h, s, v, = img.unbind(0)
            h = h.flatten().numpy()
            s = s.flatten().numpy()
            v = v.flatten().numpy()

            rgb = []
            for h1, s1, v1 in zip(h, s, v):
                rgb.append(colorsys.hsv_to_rgb(h1, s1, v1))

            colorsys_img = torch.tensor(rgb, dtype=torch.float32)
            max_diff = (ft_img - colorsys_img).abs().max()
            self.assertLess(max_diff, 1e-5)

    def test_rgb2hsv(self):
        shape = (3, 150, 100)
        for _ in range(20):
            img = torch.rand(*shape, dtype=torch.float)
            ft_hsv_img = F_t._rgb2hsv(img).permute(1, 2, 0).flatten(0, 1)

            r, g, b, = img.unbind(0)
            r = r.flatten().numpy()
            g = g.flatten().numpy()
            b = b.flatten().numpy()

            hsv = []
            for r1, g1, b1 in zip(r, g, b):
                hsv.append(colorsys.rgb_to_hsv(r1, g1, b1))

            colorsys_img = torch.tensor(hsv, dtype=torch.float32)

123
124
125
126
127
128
129
            ft_hsv_img_h, ft_hsv_img_sv = torch.split(ft_hsv_img, [1, 2], dim=1)
            colorsys_img_h, colorsys_img_sv = torch.split(colorsys_img, [1, 2], dim=1)

            max_diff_h = ((colorsys_img_h * 2 * math.pi).sin() - (ft_hsv_img_h * 2 * math.pi).sin()).abs().max()
            max_diff_sv = (colorsys_img_sv - ft_hsv_img_sv).abs().max()
            max_diff = max(max_diff_h, max_diff_sv)

130
131
            self.assertLess(max_diff, 1e-5)

132
    def _test_adjustments(self, device):
133
134
135
136
137
138
139
        script_adjust_brightness = torch.jit.script(F_t.adjust_brightness)
        script_adjust_contrast = torch.jit.script(F_t.adjust_contrast)
        script_adjust_saturation = torch.jit.script(F_t.adjust_saturation)

        fns = ((F.adjust_brightness, F_t.adjust_brightness, script_adjust_brightness),
               (F.adjust_contrast, F_t.adjust_contrast, script_adjust_contrast),
               (F.adjust_saturation, F_t.adjust_saturation, script_adjust_saturation))
140
141
142
143
144
145
146

        for _ in range(20):
            channels = 3
            dims = torch.randint(1, 50, (2,))
            shape = (channels, dims[0], dims[1])

            if torch.randint(0, 2, (1,)) == 0:
147
                img = torch.rand(*shape, dtype=torch.float, device=device)
148
            else:
149
                img = torch.randint(0, 256, shape, dtype=torch.uint8, device=device)
150

151
            factor = 3 * torch.rand(1).item()
152
            img_clone = img.clone()
153
            for f, ft, sft in fns:
154

155
156
                ft_img = ft(img, factor).cpu()
                sft_img = sft(img, factor).cpu()
157
158
                if not img.dtype.is_floating_point:
                    ft_img = ft_img.to(torch.float) / 255
159
                    sft_img = sft_img.to(torch.float) / 255
160
161
162
163
164
165
166
167

                img_pil = transforms.ToPILImage()(img)
                f_img_pil = f(img_pil, factor)
                f_img = transforms.ToTensor()(f_img_pil)

                # F uses uint8 and F_t uses float, so there is a small
                # difference in values caused by (at most 5) truncations.
                max_diff = (ft_img - f_img).abs().max()
168
                max_diff_scripted = (sft_img - f_img).abs().max()
169
                self.assertLess(max_diff, 5 / 255 + 1e-5)
170
                self.assertLess(max_diff_scripted, 5 / 255 + 1e-5)
171
                self.assertTrue(torch.equal(img, img_clone))
172

173
            # test for class interface
174
            f = transforms.ColorJitter(brightness=factor)
175
176
177
            scripted_fn = torch.jit.script(f)
            scripted_fn(img)

178
            f = transforms.ColorJitter(contrast=factor)
179
180
181
            scripted_fn = torch.jit.script(f)
            scripted_fn(img)

182
            f = transforms.ColorJitter(saturation=factor)
183
184
185
186
187
188
189
            scripted_fn = torch.jit.script(f)
            scripted_fn(img)

        f = transforms.ColorJitter(brightness=1)
        scripted_fn = torch.jit.script(f)
        scripted_fn(img)

190
191
192
193
194
195
196
    def test_adjustments(self):
        self._test_adjustments("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_adjustments_cuda(self):
        self._test_adjustments("cuda")

197
    def test_rgb_to_grayscale(self):
198
        script_rgb_to_grayscale = torch.jit.script(F_t.rgb_to_grayscale)
199
        img_tensor = torch.randint(0, 255, (3, 16, 16), dtype=torch.uint8)
200
        img_tensor_clone = img_tensor.clone()
201
202
203
204
        grayscale_tensor = F_t.rgb_to_grayscale(img_tensor).to(int)
        grayscale_pil_img = torch.tensor(np.array(F.to_grayscale(F.to_pil_image(img_tensor)))).to(int)
        max_diff = (grayscale_tensor - grayscale_pil_img).abs().max()
        self.assertLess(max_diff, 1.0001)
205
        self.assertTrue(torch.equal(img_tensor, img_tensor_clone))
206
207
208
        # scriptable function test
        grayscale_script = script_rgb_to_grayscale(img_tensor).to(int)
        self.assertTrue(torch.equal(grayscale_script, grayscale_tensor))
209

210
    def _test_center_crop(self, device):
211
212
        script_center_crop = torch.jit.script(F.center_crop)

213
        img_tensor, pil_img = self._create_data(32, 34, device=device)
214
215
216
217
218
219
220
221

        cropped_pil_image = F.center_crop(pil_img, [10, 11])

        cropped_tensor = F.center_crop(img_tensor, [10, 11])
        self.compareTensorToPIL(cropped_tensor, cropped_pil_image)

        cropped_tensor = script_center_crop(img_tensor, [10, 11])
        self.compareTensorToPIL(cropped_tensor, cropped_pil_image)
222

223
224
225
226
227
228
229
230
    def test_center_crop(self):
        self._test_center_crop("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_center_crop_cuda(self):
        self._test_center_crop("cuda")

    def _test_five_crop(self, device):
231
232
        script_five_crop = torch.jit.script(F.five_crop)

233
        img_tensor, pil_img = self._create_data(32, 34, device=device)
234
235
236
237
238
239
240
241
242
243

        cropped_pil_images = F.five_crop(pil_img, [10, 11])

        cropped_tensors = F.five_crop(img_tensor, [10, 11])
        for i in range(5):
            self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])

        cropped_tensors = script_five_crop(img_tensor, [10, 11])
        for i in range(5):
            self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])
244

245
246
247
248
249
250
251
252
    def test_five_crop(self):
        self._test_five_crop("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_five_crop_cuda(self):
        self._test_five_crop("cuda")

    def _test_ten_crop(self, device):
253
254
        script_ten_crop = torch.jit.script(F.ten_crop)

255
        img_tensor, pil_img = self._create_data(32, 34, device=device)
256
257
258
259
260
261
262
263
264
265

        cropped_pil_images = F.ten_crop(pil_img, [10, 11])

        cropped_tensors = F.ten_crop(img_tensor, [10, 11])
        for i in range(10):
            self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])

        cropped_tensors = script_ten_crop(img_tensor, [10, 11])
        for i in range(10):
            self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])
266

267
268
269
270
271
272
273
274
    def test_ten_crop(self):
        self._test_ten_crop("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_ten_crop_cuda(self):
        self._test_ten_crop("cuda")

    def _test_pad(self, device):
275
        script_fn = torch.jit.script(F_t.pad)
276
        tensor, pil_img = self._create_data(7, 8, device=device)
277
278
279
280
281
282
283
284
285
286
287
288

        for dt in [None, torch.float32, torch.float64]:
            if dt is not None:
                # This is a trivial cast to float of uint8 data to test all cases
                tensor = tensor.to(dt)
            for pad in [2, [3, ], [0, 3], (3, 3), [4, 2, 4, 3]]:
                configs = [
                    {"padding_mode": "constant", "fill": 0},
                    {"padding_mode": "constant", "fill": 10},
                    {"padding_mode": "constant", "fill": 20},
                    {"padding_mode": "edge"},
                    {"padding_mode": "reflect"},
289
                    {"padding_mode": "symmetric"},
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
                ]
                for kwargs in configs:
                    pad_tensor = F_t.pad(tensor, pad, **kwargs)
                    pad_pil_img = F_pil.pad(pil_img, pad, **kwargs)

                    pad_tensor_8b = pad_tensor
                    # we need to cast to uint8 to compare with PIL image
                    if pad_tensor_8b.dtype != torch.uint8:
                        pad_tensor_8b = pad_tensor_8b.to(torch.uint8)

                    self.compareTensorToPIL(pad_tensor_8b, pad_pil_img, msg="{}, {}".format(pad, kwargs))

                    if isinstance(pad, int):
                        script_pad = [pad, ]
                    else:
                        script_pad = pad
                    pad_tensor_script = script_fn(tensor, script_pad, **kwargs)
                    self.assertTrue(pad_tensor.equal(pad_tensor_script), msg="{}, {}".format(pad, kwargs))
308

309
310
311
        with self.assertRaises(ValueError, msg="Padding can not be negative for symmetric padding_mode"):
            F_t.pad(tensor, (-2, -3), padding_mode="symmetric")

312
313
314
315
316
317
318
319
320
321
    def test_pad_cpu(self):
        self._test_pad("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_pad_cuda(self):
        self._test_pad("cuda")

    def _test_adjust_gamma(self, device):
        script_fn = torch.jit.script(F.adjust_gamma)
        tensor, pil_img = self._create_data(26, 36, device=device)
322
323
324
325
326
327
328
329
330
331

        for dt in [torch.float64, torch.float32, None]:

            if dt is not None:
                tensor = F.convert_image_dtype(tensor, dt)

            gammas = [0.8, 1.0, 1.2]
            gains = [0.7, 1.0, 1.3]
            for gamma, gain in zip(gammas, gains):

332
333
                adjusted_tensor = F.adjust_gamma(tensor, gamma, gain)
                adjusted_pil = F.adjust_gamma(pil_img, gamma, gain)
334
335
336
337
338
339
340
341
342
343
                scripted_result = script_fn(tensor, gamma, gain)
                self.assertEqual(adjusted_tensor.dtype, scripted_result.dtype)
                self.assertEqual(adjusted_tensor.size()[1:], adjusted_pil.size[::-1])

                rbg_tensor = adjusted_tensor
                if adjusted_tensor.dtype != torch.uint8:
                    rbg_tensor = F.convert_image_dtype(adjusted_tensor, torch.uint8)

                self.compareTensorToPIL(rbg_tensor, adjusted_pil)

344
345
346
347
348
349
350
351
                self.assertTrue(adjusted_tensor.allclose(scripted_result))

    def test_adjust_gamma_cpu(self):
        self._test_adjust_gamma("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_adjust_gamma_cuda(self):
        self._test_adjust_gamma("cuda")
352

353
    def _test_resize(self, device):
vfdev's avatar
vfdev committed
354
        script_fn = torch.jit.script(F_t.resize)
355
        tensor, pil_img = self._create_data(26, 36, device=device)
vfdev's avatar
vfdev committed
356
357
358
359
360

        for dt in [None, torch.float32, torch.float64]:
            if dt is not None:
                # This is a trivial cast to float of uint8 data to test all cases
                tensor = tensor.to(dt)
361
            for size in [32, 26, [32, ], [32, 32], (32, 32), [26, 35]]:
vfdev's avatar
vfdev committed
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
                for interpolation in [BILINEAR, BICUBIC, NEAREST]:
                    resized_tensor = F_t.resize(tensor, size=size, interpolation=interpolation)
                    resized_pil_img = F_pil.resize(pil_img, size=size, interpolation=interpolation)

                    self.assertEqual(
                        resized_tensor.size()[1:], resized_pil_img.size[::-1], msg="{}, {}".format(size, interpolation)
                    )

                    if interpolation != NEAREST:
                        # We can not check values if mode = NEAREST, as results are different
                        # E.g. resized_tensor  = [[a, a, b, c, d, d, e, ...]]
                        # E.g. resized_pil_img = [[a, b, c, c, d, e, f, ...]]
                        resized_tensor_f = resized_tensor
                        # we need to cast to uint8 to compare with PIL image
                        if resized_tensor_f.dtype == torch.uint8:
                            resized_tensor_f = resized_tensor_f.to(torch.float)

                        # Pay attention to high tolerance for MAE
                        self.approxEqualTensorToPIL(
                            resized_tensor_f, resized_pil_img, tol=8.0, msg="{}, {}".format(size, interpolation)
                        )

                    if isinstance(size, int):
                        script_size = [size, ]
                    else:
                        script_size = size
388
389
                    resize_result = script_fn(tensor, size=script_size, interpolation=interpolation)
                    self.assertTrue(resized_tensor.equal(resize_result), msg="{}, {}".format(size, interpolation))
vfdev's avatar
vfdev committed
390

391
392
393
394
395
396
397
398
    def test_resize_cpu(self):
        self._test_resize("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_resize_cuda(self):
        self._test_resize("cuda")

    def _test_resized_crop(self, device):
399
400
        # test values of F.resized_crop in several cases:
        # 1) resize to the same size, crop to the same size => should be identity
401
        tensor, _ = self._create_data(26, 36, device=device)
402
403
404
405
406
        for i in [0, 2, 3]:
            out_tensor = F.resized_crop(tensor, top=0, left=0, height=26, width=36, size=[26, 36], interpolation=i)
            self.assertTrue(tensor.equal(out_tensor), msg="{} vs {}".format(out_tensor[0, :5, :5], tensor[0, :5, :5]))

        # 2) resize by half and crop a TL corner
407
        tensor, _ = self._create_data(26, 36, device=device)
408
409
410
411
412
413
414
        out_tensor = F.resized_crop(tensor, top=0, left=0, height=20, width=30, size=[10, 15], interpolation=0)
        expected_out_tensor = tensor[:, :20:2, :30:2]
        self.assertTrue(
            expected_out_tensor.equal(out_tensor),
            msg="{} vs {}".format(expected_out_tensor[0, :10, :10], out_tensor[0, :10, :10])
        )

415
416
417
418
419
420
421
422
    def test_resized_crop_cpu(self):
        self._test_resized_crop("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_resized_crop_cuda(self):
        self._test_resized_crop("cuda")

    def _test_affine(self, device):
423
        # Tests on square and rectangular images
vfdev's avatar
vfdev committed
424
425
        scripted_affine = torch.jit.script(F.affine)

426
        for tensor, pil_img in [self._create_data(26, 26, device=device), self._create_data(32, 26, device=device)]:
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449

            # 1) identity map
            out_tensor = F.affine(tensor, angle=0, translate=[0, 0], scale=1.0, shear=[0.0, 0.0], resample=0)
            self.assertTrue(
                tensor.equal(out_tensor), msg="{} vs {}".format(out_tensor[0, :5, :5], tensor[0, :5, :5])
            )
            out_tensor = scripted_affine(tensor, angle=0, translate=[0, 0], scale=1.0, shear=[0.0, 0.0], resample=0)
            self.assertTrue(
                tensor.equal(out_tensor), msg="{} vs {}".format(out_tensor[0, :5, :5], tensor[0, :5, :5])
            )

            if pil_img.size[0] == pil_img.size[1]:
                # 2) Test rotation
                test_configs = [
                    (90, torch.rot90(tensor, k=1, dims=(-1, -2))),
                    (45, None),
                    (30, None),
                    (-30, None),
                    (-45, None),
                    (-90, torch.rot90(tensor, k=-1, dims=(-1, -2))),
                    (180, torch.rot90(tensor, k=2, dims=(-1, -2))),
                ]
                for a, true_tensor in test_configs:
450
451
452
453
454
455

                    out_pil_img = F.affine(
                        pil_img, angle=a, translate=[0, 0], scale=1.0, shear=[0.0, 0.0], resample=0
                    )
                    out_pil_tensor = torch.from_numpy(np.array(out_pil_img).transpose((2, 0, 1))).to(device)

456
                    for fn in [F.affine, scripted_affine]:
457
458
459
                        out_tensor = fn(
                            tensor, angle=a, translate=[0, 0], scale=1.0, shear=[0.0, 0.0], resample=0
                        )
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
                        if true_tensor is not None:
                            self.assertTrue(
                                true_tensor.equal(out_tensor),
                                msg="{}\n{} vs \n{}".format(a, out_tensor[0, :5, :5], true_tensor[0, :5, :5])
                            )
                        else:
                            true_tensor = out_tensor

                        num_diff_pixels = (true_tensor != out_pil_tensor).sum().item() / 3.0
                        ratio_diff_pixels = num_diff_pixels / true_tensor.shape[-1] / true_tensor.shape[-2]
                        # Tolerance : less than 6% of different pixels
                        self.assertLess(
                            ratio_diff_pixels,
                            0.06,
                            msg="{}\n{} vs \n{}".format(
                                ratio_diff_pixels, true_tensor[0, :7, :7], out_pil_tensor[0, :7, :7]
                            )
                        )
            else:
                test_configs = [
                    90, 45, 15, -30, -60, -120
                ]
                for a in test_configs:
483
484
485
486
487
488

                    out_pil_img = F.affine(
                        pil_img, angle=a, translate=[0, 0], scale=1.0, shear=[0.0, 0.0], resample=0
                    )
                    out_pil_tensor = torch.from_numpy(np.array(out_pil_img).transpose((2, 0, 1)))

489
                    for fn in [F.affine, scripted_affine]:
490
491
492
                        out_tensor = fn(
                            tensor, angle=a, translate=[0, 0], scale=1.0, shear=[0.0, 0.0], resample=0
                        ).cpu()
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509

                        num_diff_pixels = (out_tensor != out_pil_tensor).sum().item() / 3.0
                        ratio_diff_pixels = num_diff_pixels / out_tensor.shape[-1] / out_tensor.shape[-2]
                        # Tolerance : less than 3% of different pixels
                        self.assertLess(
                            ratio_diff_pixels,
                            0.03,
                            msg="{}: {}\n{} vs \n{}".format(
                                a, ratio_diff_pixels, out_tensor[0, :7, :7], out_pil_tensor[0, :7, :7]
                            )
                        )

            # 3) Test translation
            test_configs = [
                [10, 12], (-12, -13)
            ]
            for t in test_configs:
510
511
512

                out_pil_img = F.affine(pil_img, angle=0, translate=t, scale=1.0, shear=[0.0, 0.0], resample=0)

vfdev's avatar
vfdev committed
513
                for fn in [F.affine, scripted_affine]:
514
                    out_tensor = fn(tensor, angle=0, translate=t, scale=1.0, shear=[0.0, 0.0], resample=0)
515

516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
                    self.compareTensorToPIL(out_tensor, out_pil_img)

            # 3) Test rotation + translation + scale + share
            test_configs = [
                (45, [5, 6], 1.0, [0.0, 0.0]),
                (33, (5, -4), 1.0, [0.0, 0.0]),
                (45, [-5, 4], 1.2, [0.0, 0.0]),
                (33, (-4, -8), 2.0, [0.0, 0.0]),
                (85, (10, -10), 0.7, [0.0, 0.0]),
                (0, [0, 0], 1.0, [35.0, ]),
                (-25, [0, 0], 1.2, [0.0, 15.0]),
                (-45, [-10, 0], 0.7, [2.0, 5.0]),
                (-45, [-10, -10], 1.2, [4.0, 5.0]),
                (-90, [0, 0], 1.0, [0.0, 0.0]),
            ]
            for r in [0, ]:
                for a, t, s, sh in test_configs:
                    out_pil_img = F.affine(pil_img, angle=a, translate=t, scale=s, shear=sh, resample=r)
                    out_pil_tensor = torch.from_numpy(np.array(out_pil_img).transpose((2, 0, 1)))

                    for fn in [F.affine, scripted_affine]:
537
                        out_tensor = fn(tensor, angle=a, translate=t, scale=s, shear=sh, resample=r).cpu()
538
539
                        num_diff_pixels = (out_tensor != out_pil_tensor).sum().item() / 3.0
                        ratio_diff_pixels = num_diff_pixels / out_tensor.shape[-1] / out_tensor.shape[-2]
540
541
                        # Tolerance : less than 5% (cpu), 6% (cuda) of different pixels
                        tol = 0.06 if device == "cuda" else 0.05
542
543
                        self.assertLess(
                            ratio_diff_pixels,
544
                            tol,
545
546
547
                            msg="{}: {}\n{} vs \n{}".format(
                                (r, a, t, s, sh), ratio_diff_pixels, out_tensor[0, :7, :7], out_pil_tensor[0, :7, :7]
                            )
vfdev's avatar
vfdev committed
548
549
                        )

550
551
552
553
554
555
556
557
    def test_affine_cpu(self):
        self._test_affine("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_affine_cuda(self):
        self._test_affine("cuda")

    def _test_rotate(self, device):
vfdev's avatar
vfdev committed
558
559
560
        # Tests on square image
        scripted_rotate = torch.jit.script(F.rotate)

561
        for tensor, pil_img in [self._create_data(26, 26, device=device), self._create_data(32, 26, device=device)]:
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577

            img_size = pil_img.size
            centers = [
                None,
                (int(img_size[0] * 0.3), int(img_size[0] * 0.4)),
                [int(img_size[0] * 0.5), int(img_size[0] * 0.6)]
            ]

            for r in [0, ]:
                for a in range(-180, 180, 17):
                    for e in [True, False]:
                        for c in centers:

                            out_pil_img = F.rotate(pil_img, angle=a, resample=r, expand=e, center=c)
                            out_pil_tensor = torch.from_numpy(np.array(out_pil_img).transpose((2, 0, 1)))
                            for fn in [F.rotate, scripted_rotate]:
578
                                out_tensor = fn(tensor, angle=a, resample=r, expand=e, center=c).cpu()
579
580
581
582
583
584
585

                                self.assertEqual(
                                    out_tensor.shape,
                                    out_pil_tensor.shape,
                                    msg="{}: {} vs {}".format(
                                        (img_size, r, a, e, c), out_tensor.shape, out_pil_tensor.shape
                                    )
vfdev's avatar
vfdev committed
586
                                )
587
588
589
590
591
592
593
594
595
596
597
598
                                num_diff_pixels = (out_tensor != out_pil_tensor).sum().item() / 3.0
                                ratio_diff_pixels = num_diff_pixels / out_tensor.shape[-1] / out_tensor.shape[-2]
                                # Tolerance : less than 2% of different pixels
                                self.assertLess(
                                    ratio_diff_pixels,
                                    0.02,
                                    msg="{}: {}\n{} vs \n{}".format(
                                        (img_size, r, a, e, c),
                                        ratio_diff_pixels,
                                        out_tensor[0, :7, :7],
                                        out_pil_tensor[0, :7, :7]
                                    )
vfdev's avatar
vfdev committed
599
600
                                )

601
602
603
604
605
606
607
608
    def test_rotate_cpu(self):
        self._test_rotate("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_rotate_cuda(self):
        self._test_rotate("cuda")

    def _test_perspective(self, device):
609
610
611

        from torchvision.transforms import RandomPerspective

612
        for tensor, pil_img in [self._create_data(26, 34, device=device), self._create_data(26, 26, device=device)]:
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631

            scripted_tranform = torch.jit.script(F.perspective)

            test_configs = [
                [[[0, 0], [33, 0], [33, 25], [0, 25]], [[3, 2], [32, 3], [30, 24], [2, 25]]],
                [[[3, 2], [32, 3], [30, 24], [2, 25]], [[0, 0], [33, 0], [33, 25], [0, 25]]],
                [[[3, 2], [32, 3], [30, 24], [2, 25]], [[5, 5], [30, 3], [33, 19], [4, 25]]],
            ]
            n = 10
            test_configs += [
                RandomPerspective.get_params(pil_img.size[0], pil_img.size[1], i / n) for i in range(n)
            ]

            for r in [0, ]:
                for spoints, epoints in test_configs:
                    out_pil_img = F.perspective(pil_img, startpoints=spoints, endpoints=epoints, interpolation=r)
                    out_pil_tensor = torch.from_numpy(np.array(out_pil_img).transpose((2, 0, 1)))

                    for fn in [F.perspective, scripted_tranform]:
632
                        out_tensor = fn(tensor, startpoints=spoints, endpoints=epoints, interpolation=r).cpu()
633
634
635

                        num_diff_pixels = (out_tensor != out_pil_tensor).sum().item() / 3.0
                        ratio_diff_pixels = num_diff_pixels / out_tensor.shape[-1] / out_tensor.shape[-2]
636
                        # Tolerance : less than 5% of different pixels
637
638
                        self.assertLess(
                            ratio_diff_pixels,
639
                            0.05,
640
641
642
643
644
645
646
647
                            msg="{}: {}\n{} vs \n{}".format(
                                (r, spoints, epoints),
                                ratio_diff_pixels,
                                out_tensor[0, :7, :7],
                                out_pil_tensor[0, :7, :7]
                            )
                        )

648
649
650
651
652
653
654
    def test_perspective_cpu(self):
        self._test_perspective("cpu")

    @unittest.skipIf(not torch.cuda.is_available(), reason="Skip if no CUDA device")
    def test_perspective_cuda(self):
        self._test_perspective("cuda")

655
656
657

if __name__ == '__main__':
    unittest.main()