test_image.py 17.4 KB
Newer Older
1
import glob
2
3
import io
import os
4
5
import sys
from pathlib import Path
6

7
import pytest
8
import numpy as np
9
import torch
10
from PIL import Image, __version__ as PILLOW_VERSION
11
import torchvision.transforms.functional as F
12
from common_utils import needs_cuda, assert_equal
13

14
from torchvision.io.image import (
15
    decode_png, decode_jpeg, encode_jpeg, write_jpeg, decode_image, read_file,
16
    encode_png, write_png, write_file, ImageReadMode, read_image)
Francisco Massa's avatar
Francisco Massa committed
17

18
IMAGE_ROOT = os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")
19
20
FAKEDATA_DIR = os.path.join(IMAGE_ROOT, "fakedata")
IMAGE_DIR = os.path.join(FAKEDATA_DIR, "imagefolder")
21
DAMAGED_JPEG = os.path.join(IMAGE_ROOT, 'damaged_jpeg')
22
ENCODE_JPEG = os.path.join(IMAGE_ROOT, "encode_jpeg")
23
INTERLACED_PNG = os.path.join(IMAGE_ROOT, "interlaced_png")
24
IS_WINDOWS = sys.platform in ('win32', 'cygwin')
25
PILLOW_VERSION = tuple(int(x) for x in PILLOW_VERSION.split('.'))
26
27
28
29
30
31
32
33


def _get_safe_image_name(name):
    # Used when we need to change the pytest "id" for an "image path" parameter.
    # If we don't, the test id (i.e. its name) will contain the whole path to the image, which is machine-specific,
    # and this creates issues when the test is running in a different machine than where it was collected
    # (typically, in fb internal infra)
    return name.split(os.path.sep)[-1]
34
35
36
37


def get_images(directory, img_ext):
    assert os.path.isdir(directory)
38
39
40
41
    image_paths = glob.glob(directory + f'/**/*{img_ext}', recursive=True)
    for path in image_paths:
        if path.split(os.sep)[-2] not in ['damaged_jpeg', 'jpeg_write']:
            yield path
42
43


44
45
46
47
48
49
50
51
52
53
54
55
56
def pil_read_image(img_path):
    with Image.open(img_path) as img:
        return torch.from_numpy(np.array(img))


def normalize_dimensions(img_pil):
    if len(img_pil.shape) == 3:
        img_pil = img_pil.permute(2, 0, 1)
    else:
        img_pil = img_pil.unsqueeze(0)
    return img_pil


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
@pytest.mark.parametrize('img_path', [
    pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path))
    for jpeg_path in get_images(IMAGE_ROOT, ".jpg")
])
@pytest.mark.parametrize('pil_mode, mode', [
    (None, ImageReadMode.UNCHANGED),
    ("L", ImageReadMode.GRAY),
    ("RGB", ImageReadMode.RGB),
])
def test_decode_jpeg(img_path, pil_mode, mode):

    with Image.open(img_path) as img:
        is_cmyk = img.mode == "CMYK"
        if pil_mode is not None:
            if is_cmyk:
                # libjpeg does not support the conversion
                pytest.xfail("Decoding a CMYK jpeg isn't supported")
            img = img.convert(pil_mode)
        img_pil = torch.from_numpy(np.array(img))
        if is_cmyk:
            # flip the colors to match libjpeg
            img_pil = 255 - img_pil

    img_pil = normalize_dimensions(img_pil)
    data = read_file(img_path)
    img_ljpeg = decode_image(data, mode=mode)

    # Permit a small variation on pixel values to account for implementation
    # differences between Pillow and LibJPEG.
    abs_mean_diff = (img_ljpeg.type(torch.float32) - img_pil).abs().mean().item()
    assert abs_mean_diff < 2


def test_decode_jpeg_errors():
    with pytest.raises(RuntimeError, match="Expected a non empty 1-dimensional tensor"):
        decode_jpeg(torch.empty((100, 1), dtype=torch.uint8))

    with pytest.raises(RuntimeError, match="Expected a torch.uint8 tensor"):
        decode_jpeg(torch.empty((100,), dtype=torch.float16))

    with pytest.raises(RuntimeError, match="Not a JPEG file"):
        decode_jpeg(torch.empty((100), dtype=torch.uint8))


def test_decode_bad_huffman_images():
    # sanity check: make sure we can decode the bad Huffman encoding
    bad_huff = read_file(os.path.join(DAMAGED_JPEG, 'bad_huffman.jpg'))
    decode_jpeg(bad_huff)


@pytest.mark.parametrize('img_path', [
    pytest.param(truncated_image, id=_get_safe_image_name(truncated_image))
    for truncated_image in glob.glob(os.path.join(DAMAGED_JPEG, 'corrupt*.jpg'))
])
def test_damaged_corrupt_images(img_path):
    # Truncated images should raise an exception
    data = read_file(img_path)
    if 'corrupt34' in img_path:
        match_message = "Image is incomplete or truncated"
    else:
        match_message = "Unsupported marker type"
    with pytest.raises(RuntimeError, match=match_message):
        decode_jpeg(data)


@pytest.mark.parametrize('img_path', [
    pytest.param(png_path, id=_get_safe_image_name(png_path))
    for png_path in get_images(FAKEDATA_DIR, ".png")
])
@pytest.mark.parametrize('pil_mode, mode', [
    (None, ImageReadMode.UNCHANGED),
    ("L", ImageReadMode.GRAY),
    ("LA", ImageReadMode.GRAY_ALPHA),
    ("RGB", ImageReadMode.RGB),
    ("RGBA", ImageReadMode.RGB_ALPHA),
])
def test_decode_png(img_path, pil_mode, mode):

    with Image.open(img_path) as img:
        if pil_mode is not None:
            img = img.convert(pil_mode)
        img_pil = torch.from_numpy(np.array(img))

    img_pil = normalize_dimensions(img_pil)
    data = read_file(img_path)
    img_lpng = decode_image(data, mode=mode)

    tol = 0 if pil_mode is None else 1
145
146
147
148
149
150
151
152
153

    if PILLOW_VERSION >= (8, 3) and pil_mode == "LA":
        # Avoid checking the transparency channel until
        # https://github.com/python-pillow/Pillow/issues/5593#issuecomment-878244910
        # is fixed.
        # TODO: remove once fix is released in PIL. Should be > 8.3.1.
        img_lpng, img_pil = img_lpng[0], img_pil[0]

    torch.testing.assert_close(img_lpng, img_pil, atol=tol, rtol=0)
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


def test_decode_png_errors():
    with pytest.raises(RuntimeError, match="Expected a non empty 1-dimensional tensor"):
        decode_png(torch.empty((), dtype=torch.uint8))
    with pytest.raises(RuntimeError, match="Content is not png"):
        decode_png(torch.randint(3, 5, (300,), dtype=torch.uint8))


@pytest.mark.parametrize('img_path', [
    pytest.param(png_path, id=_get_safe_image_name(png_path))
    for png_path in get_images(IMAGE_DIR, ".png")
])
def test_encode_png(img_path):
    pil_image = Image.open(img_path)
    img_pil = torch.from_numpy(np.array(pil_image))
    img_pil = img_pil.permute(2, 0, 1)
    png_buf = encode_png(img_pil, compression_level=6)

    rec_img = Image.open(io.BytesIO(bytes(png_buf.tolist())))
    rec_img = torch.from_numpy(np.array(rec_img))
    rec_img = rec_img.permute(2, 0, 1)

    assert_equal(img_pil, rec_img)


def test_encode_png_errors():
    with pytest.raises(RuntimeError, match="Input tensor dtype should be uint8"):
        encode_png(torch.empty((3, 100, 100), dtype=torch.float32))

    with pytest.raises(RuntimeError, match="Compression level should be between 0 and 9"):
        encode_png(torch.empty((3, 100, 100), dtype=torch.uint8),
                   compression_level=-1)

    with pytest.raises(RuntimeError, match="Compression level should be between 0 and 9"):
        encode_png(torch.empty((3, 100, 100), dtype=torch.uint8),
                   compression_level=10)

    with pytest.raises(RuntimeError, match="The number of channels should be 1 or 3, got: 5"):
        encode_png(torch.empty((5, 100, 100), dtype=torch.uint8))


@pytest.mark.parametrize('img_path', [
    pytest.param(png_path, id=_get_safe_image_name(png_path))
    for png_path in get_images(IMAGE_DIR, ".png")
])
200
201
202
203
def test_write_png(img_path, tmpdir):
    pil_image = Image.open(img_path)
    img_pil = torch.from_numpy(np.array(pil_image))
    img_pil = img_pil.permute(2, 0, 1)
204

205
206
207
208
209
    filename, _ = os.path.splitext(os.path.basename(img_path))
    torch_png = os.path.join(tmpdir, '{0}_torch.png'.format(filename))
    write_png(img_pil, torch_png, compression_level=6)
    saved_image = torch.from_numpy(np.array(Image.open(torch_png)))
    saved_image = saved_image.permute(2, 0, 1)
210

211
    assert_equal(img_pil, saved_image)
212
213


214
215
216
217
218
def test_read_file(tmpdir):
    fname, content = 'test1.bin', b'TorchVision\211\n'
    fpath = os.path.join(tmpdir, fname)
    with open(fpath, 'wb') as f:
        f.write(content)
219

220
221
222
223
    data = read_file(fpath)
    expected = torch.tensor(list(content), dtype=torch.uint8)
    os.unlink(fpath)
    assert_equal(data, expected)
224
225
226
227
228

    with pytest.raises(RuntimeError, match="No such file or directory: 'tst'"):
        read_file('tst')


229
230
231
232
233
def test_read_file_non_ascii(tmpdir):
    fname, content = '日本語(Japanese).bin', b'TorchVision\211\n'
    fpath = os.path.join(tmpdir, fname)
    with open(fpath, 'wb') as f:
        f.write(content)
234

235
236
237
238
    data = read_file(fpath)
    expected = torch.tensor(list(content), dtype=torch.uint8)
    os.unlink(fpath)
    assert_equal(data, expected)
239
240


241
242
243
244
245
def test_write_file(tmpdir):
    fname, content = 'test1.bin', b'TorchVision\211\n'
    fpath = os.path.join(tmpdir, fname)
    content_tensor = torch.tensor(list(content), dtype=torch.uint8)
    write_file(fpath, content_tensor)
246

247
248
249
250
    with open(fpath, 'rb') as f:
        saved_content = f.read()
    os.unlink(fpath)
    assert content == saved_content
251
252


253
254
255
256
257
def test_write_file_non_ascii(tmpdir):
    fname, content = '日本語(Japanese).bin', b'TorchVision\211\n'
    fpath = os.path.join(tmpdir, fname)
    content_tensor = torch.tensor(list(content), dtype=torch.uint8)
    write_file(fpath, content_tensor)
258

259
260
261
262
    with open(fpath, 'rb') as f:
        saved_content = f.read()
    os.unlink(fpath)
    assert content == saved_content
263

264

Prabhat Roy's avatar
Prabhat Roy committed
265
266
267
268
269
@pytest.mark.parametrize('shape', [
    (27, 27),
    (60, 60),
    (105, 105),
])
270
def test_read_1_bit_png(shape, tmpdir):
271
    np_rng = np.random.RandomState(0)
272
273
274
275
276
277
278
    image_path = os.path.join(tmpdir, f'test_{shape}.png')
    pixels = np_rng.rand(*shape) > 0.5
    img = Image.fromarray(pixels)
    img.save(image_path)
    img1 = read_image(image_path)
    img2 = normalize_dimensions(torch.as_tensor(pixels * 255, dtype=torch.uint8))
    assert_equal(img1, img2)
Prabhat Roy's avatar
Prabhat Roy committed
279
280
281
282
283
284
285
286
287
288
289


@pytest.mark.parametrize('shape', [
    (27, 27),
    (60, 60),
    (105, 105),
])
@pytest.mark.parametrize('mode', [
    ImageReadMode.UNCHANGED,
    ImageReadMode.GRAY,
])
290
def test_read_1_bit_png_consistency(shape, mode, tmpdir):
291
    np_rng = np.random.RandomState(0)
292
293
294
295
296
297
298
    image_path = os.path.join(tmpdir, f'test_{shape}.png')
    pixels = np_rng.rand(*shape) > 0.5
    img = Image.fromarray(pixels)
    img.save(image_path)
    img1 = read_image(image_path, mode)
    img2 = read_image(image_path, mode)
    assert_equal(img1, img2)
Prabhat Roy's avatar
Prabhat Roy committed
299
300


301
302
303
304
305
306
307
308
309
def test_read_interlaced_png():
    imgs = list(get_images(INTERLACED_PNG, ".png"))
    with Image.open(imgs[0]) as im1, Image.open(imgs[1]) as im2:
        assert not (im1.info.get("interlace") is im2.info.get("interlace"))
    img1 = read_image(imgs[0])
    img2 = read_image(imgs[1])
    assert_equal(img1, img2)


310
@needs_cuda
311
@pytest.mark.parametrize('img_path', [
312
    pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path))
313
314
    for jpeg_path in get_images(IMAGE_ROOT, ".jpg")
])
315
316
317
318
319
@pytest.mark.parametrize('mode', [ImageReadMode.UNCHANGED, ImageReadMode.GRAY, ImageReadMode.RGB])
@pytest.mark.parametrize('scripted', (False, True))
def test_decode_jpeg_cuda(mode, img_path, scripted):
    if 'cmyk' in img_path:
        pytest.xfail("Decoding a CMYK jpeg isn't supported")
320

321
322
323
324
325
326
    data = read_file(img_path)
    img = decode_image(data, mode=mode)
    f = torch.jit.script(decode_jpeg) if scripted else decode_jpeg
    img_nvjpeg = f(data, mode=mode, device='cuda')

    # Some difference expected between jpeg implementations
327
    assert (img.float() - img_nvjpeg.cpu().float()).abs().mean() < 2
328
329
330
331
332
333


@needs_cuda
@pytest.mark.parametrize('cuda_device', ('cuda', 'cuda:0', torch.device('cuda')))
def test_decode_jpeg_cuda_device_param(cuda_device):
    """Make sure we can pass a string or a torch.device as device param"""
334
335
    path = next(path for path in get_images(IMAGE_ROOT, ".jpg") if 'cmyk' not in path)
    data = read_file(path)
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
    decode_jpeg(data, device=cuda_device)


@needs_cuda
def test_decode_jpeg_cuda_errors():
    data = read_file(next(get_images(IMAGE_ROOT, ".jpg")))
    with pytest.raises(RuntimeError, match="Expected a non empty 1-dimensional tensor"):
        decode_jpeg(data.reshape(-1, 1), device='cuda')
    with pytest.raises(RuntimeError, match="input tensor must be on CPU"):
        decode_jpeg(data.to('cuda'), device='cuda')
    with pytest.raises(RuntimeError, match="Expected a torch.uint8 tensor"):
        decode_jpeg(data.to(torch.float), device='cuda')
    with pytest.raises(RuntimeError, match="Expected a cuda device"):
        torch.ops.image.decode_jpeg_cuda(data, ImageReadMode.UNCHANGED.value, 'cpu')


352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
def test_encode_jpeg_errors():

    with pytest.raises(RuntimeError, match="Input tensor dtype should be uint8"):
        encode_jpeg(torch.empty((3, 100, 100), dtype=torch.float32))

    with pytest.raises(ValueError, match="Image quality should be a positive number "
                                         "between 1 and 100"):
        encode_jpeg(torch.empty((3, 100, 100), dtype=torch.uint8), quality=-1)

    with pytest.raises(ValueError, match="Image quality should be a positive number "
                                         "between 1 and 100"):
        encode_jpeg(torch.empty((3, 100, 100), dtype=torch.uint8), quality=101)

    with pytest.raises(RuntimeError, match="The number of channels should be 1 or 3, got: 5"):
        encode_jpeg(torch.empty((5, 100, 100), dtype=torch.uint8))

    with pytest.raises(RuntimeError, match="Input data should be a 3-dimensional tensor"):
        encode_jpeg(torch.empty((1, 3, 100, 100), dtype=torch.uint8))

    with pytest.raises(RuntimeError, match="Input data should be a 3-dimensional tensor"):
        encode_jpeg(torch.empty((100, 100), dtype=torch.uint8))


375
376
377
378
379
380
381
382
383
384
385
386
def _collect_if(cond):
    # TODO: remove this once test_encode_jpeg_reference and test_write_jpeg_reference
    # are removed
    def _inner(test_func):
        if cond:
            return test_func
        else:
            return pytest.mark.dont_collect(test_func)
    return _inner


@_collect_if(cond=IS_WINDOWS)
387
388
389
390
@pytest.mark.parametrize('img_path', [
    pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path))
    for jpeg_path in get_images(ENCODE_JPEG, ".jpg")
])
391
def test_encode_jpeg_reference(img_path):
392
    # This test is *wrong*.
393
    # It compares a torchvision-encoded jpeg with a PIL-encoded jpeg (the reference), but it
394
395
396
397
398
399
400
401
    # starts encoding the torchvision version from an image that comes from
    # decode_jpeg, which can yield different results from pil.decode (see
    # test_decode... which uses a high tolerance).
    # Instead, we should start encoding from the exact same decoded image, for a
    # valid comparison. This is done in test_encode_jpeg, but unfortunately
    # these more correct tests fail on windows (probably because of a difference
    # in libjpeg) between torchvision and PIL.
    # FIXME: make the correct tests pass on windows and remove this.
402
403
404
405
406
407
408
409
410
411
412
413
414
415
    dirname = os.path.dirname(img_path)
    filename, _ = os.path.splitext(os.path.basename(img_path))
    write_folder = os.path.join(dirname, 'jpeg_write')
    expected_file = os.path.join(
        write_folder, '{0}_pil.jpg'.format(filename))
    img = decode_jpeg(read_file(img_path))

    with open(expected_file, 'rb') as f:
        pil_bytes = f.read()
        pil_bytes = torch.as_tensor(list(pil_bytes), dtype=torch.uint8)
    for src_img in [img, img.contiguous()]:
        # PIL sets jpeg quality to 75 by default
        jpeg_bytes = encode_jpeg(src_img, quality=75)
        assert_equal(jpeg_bytes, pil_bytes)
416
417


418
@_collect_if(cond=IS_WINDOWS)
419
420
421
422
@pytest.mark.parametrize('img_path', [
    pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path))
    for jpeg_path in get_images(ENCODE_JPEG, ".jpg")
])
423
def test_write_jpeg_reference(img_path, tmpdir):
424
    # FIXME: Remove this eventually, see test_encode_jpeg_reference
425
426
    data = read_file(img_path)
    img = decode_jpeg(data)
427

428
429
430
431
432
433
    basedir = os.path.dirname(img_path)
    filename, _ = os.path.splitext(os.path.basename(img_path))
    torch_jpeg = os.path.join(
        tmpdir, '{0}_torch.jpg'.format(filename))
    pil_jpeg = os.path.join(
        basedir, 'jpeg_write', '{0}_pil.jpg'.format(filename))
434

435
    write_jpeg(img, torch_jpeg, quality=75)
436

437
438
    with open(torch_jpeg, 'rb') as f:
        torch_bytes = f.read()
439

440
441
    with open(pil_jpeg, 'rb') as f:
        pil_bytes = f.read()
442

443
    assert_equal(torch_bytes, pil_bytes)
444
445


446
447
448
@pytest.mark.skipif(IS_WINDOWS, reason=(
    'this test fails on windows because PIL uses libjpeg-turbo on windows'
))
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
@pytest.mark.parametrize('img_path', [
    pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path))
    for jpeg_path in get_images(ENCODE_JPEG, ".jpg")
])
def test_encode_jpeg(img_path):
    img = read_image(img_path)

    pil_img = F.to_pil_image(img)
    buf = io.BytesIO()
    pil_img.save(buf, format='JPEG', quality=75)

    # pytorch can't read from raw bytes so we go through numpy
    pil_bytes = np.frombuffer(buf.getvalue(), dtype=np.uint8)
    encoded_jpeg_pil = torch.as_tensor(pil_bytes)

    for src_img in [img, img.contiguous()]:
        encoded_jpeg_torch = encode_jpeg(src_img, quality=75)
        assert_equal(encoded_jpeg_torch, encoded_jpeg_pil)


469
470
471
@pytest.mark.skipif(IS_WINDOWS, reason=(
    'this test fails on windows because PIL uses libjpeg-turbo on windows'
))
472
473
474
475
@pytest.mark.parametrize('img_path', [
    pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path))
    for jpeg_path in get_images(ENCODE_JPEG, ".jpg")
])
476
477
478
479
def test_write_jpeg(img_path, tmpdir):
    tmpdir = Path(tmpdir)
    img = read_image(img_path)
    pil_img = F.to_pil_image(img)
480

481
482
    torch_jpeg = str(tmpdir / 'torch.jpg')
    pil_jpeg = str(tmpdir / 'pil.jpg')
483

484
485
    write_jpeg(img, torch_jpeg, quality=75)
    pil_img.save(pil_jpeg, quality=75)
486

487
488
    with open(torch_jpeg, 'rb') as f:
        torch_bytes = f.read()
489

490
491
    with open(pil_jpeg, 'rb') as f:
        pil_bytes = f.read()
492

493
    assert_equal(torch_bytes, pil_bytes)
494
495


496
497
if __name__ == "__main__":
    pytest.main([__file__])