-
Notifications
You must be signed in to change notification settings - Fork 1
/
loaders.py
1532 lines (1288 loc) · 66.2 KB
/
loaders.py
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
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
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
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
"""
Standardization instructions for all loaders:
- All loaders return (after the transform function) a torch tensor of shape (C, T, H, W).
- The order of augmentations is: resize, center crop, normalize. Any of them can be disabled, independently of the
others.
- The range of values *before* the normalization is [0, 1]. The mean and std values assume this range. If the
normalization is disabled, the final range is therefore [0, 1].
- The random frames and random segments are the same for all loaders, and given by self.list_random_frames and
self.list_random_segment_starts
"""
import abc
import itertools
import os
from abc import ABC, abstractmethod
from typing import Union, List, Iterator
import warnings
import PIL
import cv2
import ffmpeg
import numpy as np
import torch
import torchaudio
import torchvision
from PIL import Image
import transforms
from parameters import path_to_ffprobe, path_to_ffmpeg, parameters
loaders = parameters['loader_name']
if 'torchvideo_gulp' in loaders or 'torchvideo_pil' in loaders or 'torchvideo_video' in loaders:
import torchvideo
if 'torchvideo_gulp' in loaders:
import gulpio
from gulpio import transforms as gulp_transforms
if 'moviepy' in loaders:
import moviepy
if 'pytorchvideo_frames' in loaders or 'pytorchvideo_pyav' in loaders or 'pytorchvideo_torchvision' in loaders or \
'pytorchvideo_decord' in loaders:
from pytorchvideo.data.frame_video import FrameVideo
from pytorchvideo.data.encoded_video import EncodedVideo
import pytorchvideo.transforms
import pytorchvideo.transforms.functional
if 'pims_pyav' in loaders or 'pims_imageio' in loaders or 'pims_moviepy' in loaders:
import pims
if 'mmcv_video' in loaders or 'mmcv_image' in loaders:
import mmcv
if 'decord_video' in loaders or 'decord_video_gpu' in loaders:
import decord
if 'dali' in loaders:
from nvidia.dali import pipeline_def
import nvidia.dali.fn as fn
import nvidia.dali.types as types
import utils
from utils import NotPossibleException
mean_norm = [0.485, 0.456, 0.406]
std_norm = [0.229, 0.224, 0.225]
def get_loader(loader_name, **kwargs):
return {
'decord_video': DecordVideo,
'decord_video_gpu': DecordVideoGPU,
'pillow': Pillow,
'pillow_simd': PillowSIMD,
'opencv_image': OpenCVImage,
'opencv_video': OpenCVVideo,
'mmcv_video': MMCVVideo,
'mmcv_image': MMCVImage,
'pims_pyav': PIMSPyAV,
'pims_imageio': PIMSImageIO,
'pims_moviepy': PIMSMoviePy,
'ffmpeg': FFmpeg,
'moviepy': MoviePy,
'pytorchvideo_frames': PyTorchVideoFrames,
'pytorchvideo_pyav': PyTorchVideoPyAV,
'pytorchvideo_torchvision': PyTorchVideoTorchvision,
'pytorchvideo_decord': PyTorchVideoDecord,
'torchvision_videoloader': TorchVisionVideoReader,
'torchvision_videoloader_pyav': TorchVisionVideoReaderPyAV,
'torchvision_videoloader_cuda': TorchVisionVideoReaderCUDA,
'torchvision_readvideo': TorchVisionReadVideo,
'torchvideo_gulp': TorchVideoGULP,
'torchvideo_pil': TorchVideoPIL,
'torchvideo_video': TorchVideoVideo,
'dali': DALI,
}[loader_name](**kwargs)
class MyLoader(ABC):
can_load_audio = True
def __init__(self,
data_path,
resize=False,
short_side_size=224,
keep_aspect_ratio=True,
center_crop=False,
crop_size=224,
normalize=True,
clip_len=8,
frame_sample_rate=2,
load_format='random_frames',
list_random_frames=None,
list_random_segment_starts=None,
random_segment_duration=12,
random_segment_before_fps=False,
load_audio=False,
num_random_frames=None, # unused
num_random_segments=None, # unused
):
self.data_path = data_path
self.resize = resize
self.short_side_size = short_side_size
self.keep_aspect_ratio = keep_aspect_ratio
self.center_crop = center_crop
self.crop_size = crop_size
self.normalize = normalize
self.clip_len = clip_len
self.frame_sample_rate = frame_sample_rate
self.load_format = load_format
self.list_random_frames = list_random_frames
self.list_random_segment_starts = list_random_segment_starts
self.random_segment_duration = random_segment_duration
self.random_segment_before_fps = random_segment_before_fps
self.load_audio = load_audio
list_transforms = []
if self.resize:
if keep_aspect_ratio:
size = self.short_side_size
else:
size = (self.short_side_size, self.short_side_size)
list_transforms.append(transforms.Resize(size, interpolation='bilinear'))
if self.center_crop: # Always applied after resizing, if resizing is enabled
list_transforms.append(transforms.CenterCrop(self.crop_size))
list_transforms.append(transforms.ClipToTensor())
self.normalization_transform = transforms.Normalize(mean=mean_norm, std=std_norm)
if self.normalize:
list_transforms.append(self.normalization_transform)
self.data_transform = transforms.Compose(list_transforms)
if not self.can_load_audio and self.load_audio:
raise NotPossibleException("Cannot load audio with this loader")
@abstractmethod
def read_video(self, video_path):
pass
def transform(self, frames):
"""
This function should be implemented such that it returns a video with a CTHW format
"""
return self.data_transform(frames)
@staticmethod
def get_frames_path(video_path):
video_path = video_path.split('/')
assert video_path[-2] == 'videos'
video_path[-2] = 'frames'
video_path[-1] = video_path[-1].replace('.mp4', '')
video_path = '/'.join(video_path)
return video_path
def get_frame_indices(self, len_video, video_path=None, fps=None):
if self.load_format == 'random_frames':
frame_indices = [idx % len_video for idx in self.list_random_frames]
elif self.load_format == 'random_segments':
initial_fps = utils.get_fps(video_path) if fps is None else fps
ratio = self.frame_sample_rate / initial_fps if self.frame_sample_rate != -1 else 1
segment_starts = self.get_segment_starts(len_video, ratio)
frame_indices = [round(start + idx / ratio) for start in segment_starts for idx in
range(self.random_segment_duration)]
else: # 'all_video'
if self.frame_sample_rate == -1:
frame_indices = list(range(len_video))
else:
initial_fps = utils.get_fps(video_path) if fps is None else fps
ratio = self.frame_sample_rate / initial_fps
frame_indices = list(np.round(np.linspace(0, len_video, num=round(len_video * ratio), endpoint=False)).
astype(int))
return frame_indices
def get_segment_starts(self, len_video, ratio, return_in_original_fps=True):
"""
Returns the starting frames of segments, where the frames are given in the original video's fps.
If return_in_original_fps is True, the frames will be indexed wrt the original fps.
"""
if self.random_segment_before_fps:
max_seg_start = int(np.floor(len_video - self.random_segment_duration / ratio))
segment_starts = [idx % max_seg_start for idx in self.list_random_segment_starts]
else:
max_seg_start = round(len_video * ratio) - self.random_segment_duration
if return_in_original_fps:
segment_starts = [round((idx % max_seg_start) / ratio) for idx in self.list_random_segment_starts]
else:
segment_starts = [idx % max_seg_start for idx in self.list_random_segment_starts]
return segment_starts
def add_current_frame(self, ratio, i):
"""
When looping over all frames in a video, this function returns the number of times that frame should be added to
the final video. If the final sampling rate is larger than the original fps, the frame should be added one or
more times. If the final sampling rate is smaller than the original fps, the frame should be added zero or one
times.
"""
if self.frame_sample_rate == -1 or ratio == 1:
to_add = 1
elif ratio <= 1:
closest_frame = np.round(i * ratio)
condition = np.abs(i * ratio - closest_frame) <= np.abs((i - 1) * ratio - closest_frame) and \
np.abs(i * ratio - closest_frame) < np.abs((i + 1) * ratio - closest_frame)
to_add = 1 if condition else 0
else: # The final sampling rate is > than the original, therefore some frames are repeated
to_add = np.sum(np.round(np.array(range(int(np.ceil((i - 1) * ratio)),
int(np.ceil((i + 1) * ratio)))) / ratio) == i)
return to_add
@staticmethod
def read_audio(video_path, start_time, end_time):
"""
This function loads audio from a .wav file. It is used when the audio is not loaded by default by the loader.
The purpose of this repository is not to compare audio loaders, so this default audio loader may not be the best
one. If you want to use a different audio loader, you can override this function in your loader class.
"""
audio_path = video_path.replace('videos', 'audios').replace('.mp4', '.wav')
metadata = torchaudio.info(audio_path)
sample_rate = metadata.sample_rate
frame_offset = round(sample_rate * start_time)
if end_time is None:
num_frames = -1
else:
num_frames = round(sample_rate * (end_time - start_time))
waveform, sample_rate = torchaudio.load(audio_path, frame_offset=frame_offset, num_frames=num_frames)
return waveform
def get_audio_segment(self, video_path, start, ratio, initial_fps=None):
"""
Convenient function to load audio when we only have the start in frames (not in seconds). It deals with the
conversion between fps and frames, and finds the corresponding audio segment.
"""
initial_fps = utils.get_fps(video_path) if initial_fps is None else initial_fps
if self.random_segment_before_fps:
start_time = start / initial_fps
else:
start_time = start / initial_fps * ratio
duration_time = self.random_segment_duration / initial_fps / ratio
try:
audio_segment = self.read_audio(video_path, start_time, start_time + duration_time)
except RuntimeError:
# Sometimes the video file is problematic, and the loaders return an incorrect number of frames
start_time = start_time / 2
audio_segment = self.read_audio(video_path, start_time, start_time + duration_time)
return audio_segment
class DecordVideo(MyLoader):
"""https://github.com/dmlc/decord"""
device = 'cpu'
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# Only resize has been done as part of read_video
self.data_transform = transforms.Compose([t for t in self.data_transform.transforms
if not isinstance(t, transforms.Resize)])
def read_video(self, video_path):
"""Load video content using Decord"""
assert os.path.getsize(video_path) > 1 * 1024, 'Hanging issue'
reader = decord.VideoReader if not self.load_audio else decord.AVReader
ctx = {'cpu': decord.cpu(0), 'gpu': decord.gpu(0)}[self.device]
if not self.resize:
vr = reader(video_path, num_threads=1, ctx=ctx)
else:
if self.keep_aspect_ratio:
# Find the width and height of the video
new_width, new_height = utils.get_proportional_sizes(self.short_side_size, video_path)
else:
new_width, new_height = self.short_side_size, self.short_side_size
vr = reader(video_path, width=new_width, height=new_height, num_threads=1, ctx=ctx)
decord.bridge.set_bridge('torch')
len_video = len(vr)
fps = (vr._AVReader__video_reader if self.load_audio else vr).get_avg_fps()
frame_indices = self.get_frame_indices(len_video, video_path, fps)
buffer = vr.get_batch(frame_indices)
# If self.load_audio, the audio parameter is a list of audios associated to each retrieved frame
audio, video = buffer if self.load_audio else (None, buffer)
if self.load_audio:
audio = torch.cat(audio, dim=1) # Concatenate the audios for all the frames
return video, audio
def transform(self, frames):
frames = self.data_transform(frames)
return frames
class DecordVideoGPU(DecordVideo):
"""
Note that only CPU versions are provided with PYPI now. Please build from source to enable GPU accelerator.
https://github.com/dmlc/decord
"""
device = 'gpu'
class Pillow(MyLoader):
"""
Load from pre-extracted frames.
The default transforms already consider the case where the frames are PIL images, using PIL transforms in that case.
"""
can_load_audio = False
def __init__(self, *args, **kwargs):
assert "post" not in PIL.__version__, \
"Pillow-SIMD is installed instead of Pillow. Specify the loader pillow_simd instead of pillow. " \
"If you want to use Pillow, uninstall Pillow-SIMD and install Pillow."
super().__init__(*args, **kwargs)
def read_video(self, video_path):
frames_path = self.get_frames_path(video_path)
len_video = len(os.listdir(frames_path)) # Alternatively, read metadata for the len
frame_indices = self.get_frame_indices(len_video, video_path)
video = [Image.open(os.path.join(frames_path, '%07d.png' % idx)).convert('RGB') for idx in frame_indices]
return video, None
class PillowSIMD(Pillow):
"""
Load from pre-extracted frames.
Fork from PIL: https://github.com/uploadcare/pillow-simd
Pillow and Pillow-SIMD are not compatible. You can only have one of the installed. We could have implemented the two
of them using the same class, and the one installed would be used. However, we decided to keep them separate
explicitly so that the user knows which one is being used.
"""
def __init__(self, *args, **kwargs):
assert "post" in PIL.__version__, \
"Pillow is installed instead of Pillow-SIMD. Specify the loader `pillow' instead of `pillow_simd'. " \
"If you want to use Pillow-SIMD, uninstall Pillow and install Pillow-SIMD."
super().__init__(*args, **kwargs)
class OpenCVImage(MyLoader):
"""Load from pre-extracted frames."""
can_load_audio = False
def read_video(self, video_path):
frames_path = self.get_frames_path(video_path)
len_video = len(os.listdir(frames_path)) # Alternatively, read metadata for the len
frame_indices = self.get_frame_indices(len_video, video_path)
# The ::-1 is to convert from BGR to RGB
video = [cv2.imread(os.path.join(frames_path, '%07d.png' % idx))[:, :, ::-1] for idx in frame_indices]
return video, None
class MMCVVideo(MyLoader):
"""MMCV does not have audio support."""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def read_video(self, video_path):
video_reader = mmcv.VideoReader(video_path)
len_video = video_reader.frame_cnt
if self.load_format == 'random_frames':
frame_indices = self.get_frame_indices(len_video)
video = [video_reader[idx] for idx in frame_indices]
audio = None
if self.load_audio:
raise NotPossibleException('The default audio reader does not support random frames')
elif self.load_format == 'random_segments':
initial_fps = video_reader.fps
ratio = self.frame_sample_rate / initial_fps if self.frame_sample_rate != -1 else 1
segment_starts = self.get_segment_starts(len_video, ratio)
video = []
audio = [] if self.load_audio else None
for start in segment_starts:
video_segment = video_reader[start:start + round(self.random_segment_duration / ratio)]
if self.frame_sample_rate != -1:
indices_resample = np.round(np.linspace(0, round(self.random_segment_duration / ratio),
num=round(self.random_segment_duration),
endpoint=False)).astype(int)
video_segment = [video_segment[idx] for idx in indices_resample]
video.extend(video_segment)
if self.load_audio: # Use default implementation, no audio support
audio_segment = self.get_audio_segment(video_path, start, ratio, initial_fps)
audio.append(audio_segment)
else: # load_format == 'all_video'
video = video_reader[:]
# For some reason, video_reader.frame_cnt sometimes returns more frames than the correct number
video = [v for v in video if v is not None]
audio = None
if self.load_audio: # Use default implementation, no audio support
audio = self.read_audio(video_path, 0, None)
# Convert from BGR to RGB
video = [v[..., ::-1] for v in video]
"""
mmcv has some extra functionalities like the following:
# obtain basic information
print(len(video))
print(video.width, video.height, video.resolution, video.fps)
# iterate over all frames
for frame in video:
print(frame.shape)
# read the next frame
img = video.read()
# cut a video clip
mmcv.cut_video('test.mp4', 'clip1.mp4', start=3, end=10, vcodec='h264')
# resize a video with the specified size
mmcv.resize_video('test.mp4', 'resized1.mp4', (360, 240))
The last two these do not return a video, just save a new video, so they are not useful in a data loading
context. In the background, they use ffmpeg.
"""
return video, audio
class MMCVImage(MyLoader):
can_load_audio = False
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def read_video(self, video_path):
frames_path = self.get_frames_path(video_path)
len_video = len(os.listdir(frames_path)) # Alternatively, read metadata for the len
frame_indices = self.get_frame_indices(len_video, video_path)
# The ::-1 is to convert from BGR to RGB
video = [mmcv.imread(os.path.join(frames_path, '%07d.png' % idx))[..., ::-1] for idx in frame_indices]
return video, None
def transform(self, frames):
transformed_frames = []
for frame in frames:
if self.resize:
if self.keep_aspect_ratio:
new_h, new_w = transforms.get_resize_sizes(*frame.shape[:2], self.short_side_size)
else:
new_h, new_w = self.short_side_size, self.short_side_size
frame = mmcv.imresize(frame, (new_w, new_h))
if self.center_crop:
im_h, im_w = frame.shape[:2]
h = w = self.crop_size
if self.crop_size > im_w or self.crop_size > im_h:
error_msg = (
'Initial image size should be larger then '
'cropped size but got cropped sizes : ({w}, {h}) while '
'initial image is ({im_w}, {im_h})'.format(
im_w=im_w, im_h=im_h, w=w, h=h))
raise ValueError(error_msg)
x1 = int(round((im_w - w) / 2.))
y1 = int(round((im_h - h) / 2.))
bboxes = np.array([x1, y1, x1 + self.crop_size, y1 + self.crop_size])
frame = mmcv.imcrop(frame, bboxes) # Note that this allows to crop multiple patches at once
transformed_frames.append(frame)
transformed_frames = np.stack(transformed_frames) # (T, H, W, C)
transformed_frames = transformed_frames.transpose([3, 0, 1, 2]) # (C, T, H, W)
transformed_frames = transformed_frames / 255.
transformed_frames = torch.tensor(transformed_frames)
if self.normalize:
transformed_frames = self.normalization_transform(transformed_frames)
return transformed_frames
class OpenCVVideo(MyLoader):
def read_video(self, video_path):
# Open the video file
cap = cv2.VideoCapture(video_path)
# Get the frame rate of the video
fps = cap.get(cv2.CAP_PROP_FPS)
ratio = self.frame_sample_rate / fps if self.frame_sample_rate != -1 else 1
if self.load_format == 'random_frames':
len_video = utils.get_duration(video_path)
frame_indices = self.get_frame_indices(len_video)
# There are two ways of loading frames:
"""
# 1) iterating over the video
# Probably very inefficient unless the frame density is very high
cap.set(cv2.CAP_PROP_POS_FRAMES, 0) # Start at frame 0
video = []
ret = True
i = 0
while ret:
ret, frame = cap.read()
if ret and i in frame_indices:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
frame = Image.fromarray(frame)
video.append(frame)
i += 1
"""
# 2) using cap.set(cv2.CAP_PROP_POS_FRAMES, frame_index)
video = []
for frame_index in frame_indices:
cap.set(cv2.CAP_PROP_POS_FRAMES, frame_index)
ret, frame = cap.read()
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
frame = Image.fromarray(frame)
video.append(frame)
audio = None
if self.load_audio:
raise NotPossibleException('OpenCVVideo does not support loading audio from random frames')
elif self.load_format == 'random_segments':
len_video = utils.get_duration(video_path)
segment_starts = self.get_segment_starts(len_video, ratio)
video = []
audio = [] if self.load_audio else None
for start in segment_starts:
cap.set(cv2.CAP_PROP_POS_FRAMES, start)
for i in range(start, round(start + self.random_segment_duration / ratio)):
video, ret = self.obtain_frame(cap, video, ratio, i)
if self.load_audio:
audio_segment = self.get_audio_segment(video_path, start, ratio)
audio.append(audio_segment)
else: # load_format == 'all_video'
cap.set(cv2.CAP_PROP_POS_FRAMES, 0) # Start at frame 0
video = []
ret = True
i = 0
while ret:
video, ret = self.obtain_frame(cap, video, ratio, i)
i += 1
audio = None
if self.load_audio:
audio = self.read_audio(video_path, 0, None)
# Release the video capture object
cap.release()
return video, audio
def obtain_frame(self, cap, video, ratio, i):
ret, frame = cap.read()
to_add = self.add_current_frame(ratio, i)
if ret and to_add > 0:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
frame = Image.fromarray(frame)
video.extend([frame] * to_add)
return video, ret
class PIMS(MyLoader, abc.ABC):
"""
http://soft-matter.github.io/pims/dev/video.html
"""
reader = None
def __init__(self, **kwargs):
super().__init__(**kwargs)
def read_video(self, video_path):
reader = {'video': pims.Video, # Same as pims.PyAVVideoReader or pims.PyAVReaderTimed
'imageio': pims.ImageIOReader,
'moviepy': pims.MoviePyReader}[self.reader]
video_reader = reader(video_path)
len_video = len(video_reader)
frame_rate = video_reader.frame_rate
ratio = self.frame_sample_rate / frame_rate if self.frame_sample_rate != -1 else 1
if self.load_format == 'random_frames':
frame_indices = self.get_frame_indices(len_video)
video = [video_reader.get_frame(idx) for idx in frame_indices]
audio = None
if self.load_audio:
raise NotPossibleException('PIMS does not support loading audio from random frames')
elif self.load_format == 'random_segments':
segment_starts = self.get_segment_starts(len_video, ratio)
video = []
audio = [] if self.load_audio else None
for start in segment_starts:
video_segment = []
for i in range(start, start + round(self.random_segment_duration / ratio)):
video_segment, _ = self.obtain_frame(video_reader, video_segment, ratio, i)
video.append(video_segment)
if self.load_audio:
audio_segment = self.get_audio_segment(video_path, start, ratio)
audio.append(audio_segment)
video = np.concatenate(video)
else: # load_format == 'all_video'
video = []
for i in range(len_video):
video, end_video = self.obtain_frame(video_reader, video, ratio, i)
if end_video:
break
audio = None
if self.load_audio:
audio = self.read_audio(video_path, 0, None)
return video, audio
def obtain_frame(self, video_reader, video, ratio, i):
# The different get_frame calls will not be independent, because there is some caching going on.
# Sequential calls (or calls to frames close to the last frame) to get_frame will be faster than random calls.
# If the first frame is very well into the video, the first call to get_frame will probably be slow
to_add = self.add_current_frame(ratio, i)
if to_add > 0:
try:
frame = video_reader.get_frame(i)
except IndexError:
"""Sometimes, when the reader is imageio, depending on the formatting of the video, the duration of the
video is not read properly (similarly to mmcv reader)"""
return video, True
video.extend([frame] * to_add)
return video, False
class PIMSPyAV(PIMS):
reader = 'video'
class PIMSImageIO(PIMS):
"""Slower than using AV. Implements interface with ffmpeg through a Pipe."""
reader = 'imageio'
class PIMSMoviePy(PIMS):
"""Slower than using AV. Implements interface with ffmpeg through a Pipe."""
reader = 'moviepy'
class FFmpeg(MyLoader):
def __init__(self, **kwargs):
super().__init__(**kwargs)
def read_video(self, video_path):
len_video = utils.get_duration(video_path, method='ffprobe')
frame_rate = utils.get_fps(video_path, method='ffprobe')
ratio = self.frame_sample_rate / frame_rate if self.frame_sample_rate != -1 else 1
if self.load_format == 'random_frames':
frame_indices = self.get_frame_indices(len_video)
fps, width, height = self.get_metadata(video_path)
video = [self.read_frame_time(video_path, idx, fps, width, height) for idx in frame_indices]
video = np.transpose(np.stack(video), (3, 0, 1, 2))
audio = None
if self.load_audio:
raise NotPossibleException('FFmpeg does not support loading audio from random frames')
elif self.load_format == 'random_segments':
segment_starts = self.get_segment_starts(len_video, ratio, return_in_original_fps=False)
video = []
audio = [] if self.load_audio else None
for start in segment_starts:
duration_frames = self.random_segment_duration / (ratio if self.random_segment_before_fps else 1)
video_seg, audio_seg = self.read_video_start_end(video_path, start, start + duration_frames,
self.random_segment_before_fps)
video.append(video_seg)
if self.load_audio:
audio.append(audio_seg)
video = torch.cat(video, dim=1)
else: # load_format == 'all_video'
video, audio = self.read_video_start_end(video_path, 0, None)
return video, audio
def ffmpeg_transforms(self, v, width, height):
if self.resize:
if self.keep_aspect_ratio:
v = ffmpeg.filter(v, filter_name='scale',
w=f"if(gt(iw,ih),-1,{self.short_side_size})",
h=f"if(gt(iw,ih),{self.short_side_size},-1)")
rescale_factor = self.short_side_size / min(width, height)
width = round(width * rescale_factor)
height = round(height * rescale_factor)
else:
v = ffmpeg.filter(v, filter_name='scale', w=self.short_side_size, h=self.short_side_size)
width = height = self.short_side_size
if self.center_crop:
aw, ah = 0.5, 0.5 # Center crop
v = ffmpeg.crop(v,
'(iw - {})*{}'.format(self.crop_size, aw),
'(ih - {})*{}'.format(self.crop_size, ah),
str(self.crop_size),
str(self.crop_size))
width = height = self.crop_size
return v, width, height
def read_video_start_end(self, video_path, start_frame, end_frame, frame_before_fps=True):
# If frame_before_fps is True, the start and end frames are given in the original frame rate.
# Get width and height to reshape later, as well as fps
fps, width, height = self.get_metadata(video_path)
if not frame_before_fps:
fps = self.frame_sample_rate
start_seconds = start_frame / fps if start_frame is not None else 0
end_seconds = end_frame / fps if end_frame is not None else None
if end_frame is None:
v = ffmpeg.input(video_path, ss=start_seconds)
if self.frame_sample_rate != -1:
v = ffmpeg.filter(v, filter_name='fps', fps=self.frame_sample_rate)
else:
v = ffmpeg.input(video_path)
"""
This commented out code is equivalent to trimming the video later, but in seconds instead of frame idx.
I implement it with trim because it is more flexible: it allows to select the time before and after changing
fps. The problem with trim is that then the audio cannot be extracted directly from v.
v_initial = ffmpeg.input(video_path, ss=start_seconds, t=end_seconds - start_seconds)
"""
if frame_before_fps:
v = v.trim(start_frame=start_frame, end_frame=end_frame)
if self.frame_sample_rate != -1:
v = ffmpeg.filter(v, filter_name='fps', fps=self.frame_sample_rate)
else:
if self.frame_sample_rate != -1:
v = ffmpeg.filter(v, filter_name='fps', fps=self.frame_sample_rate)
v = v.trim(start_frame=start_frame, end_frame=end_frame) # This accounts for the fps change already
# For some reason this is necessary to do the trimming.
# See https://github.com/kkroening/ffmpeg-python/issues/184
v = v.setpts('PTS-STARTPTS')
v, width, height = self.ffmpeg_transforms(v, width, height)
out, _ = (
v.output('pipe:', format='rawvideo', pix_fmt='rgb24').
run(capture_stdout=True, quiet=True, cmd=path_to_ffmpeg)
)
video = np.frombuffer(out, np.uint8)
video = video.reshape([-1, height, width, 3])
video = torch.from_numpy(np.array(video))
video = video.permute(3, 0, 1, 2) # [C, T, H, W]
audio = None
if self.load_audio:
v = ffmpeg.input(video_path, ss=start_seconds) if end_seconds is None else \
ffmpeg.input(video_path, ss=start_seconds, t=end_seconds - start_seconds)
a = v.audio
out_audio, _ = (
a.output('pipe:', format='s16le', acodec='pcm_s16le', ac=1). # ar=sample_rate).
run(capture_stdout=True, quiet=True, cmd=path_to_ffmpeg)
)
audio = np.frombuffer(out_audio, np.int16)
audio = torch.from_numpy(np.array(audio))
return video, audio
@staticmethod
def get_metadata(video_path):
probe = ffmpeg.probe(video_path, cmd=path_to_ffprobe)
video_info = next(s for s in probe['streams'] if s['codec_type'] == 'video')
width = video_info['width']
height = video_info['height']
fps_num, fps_den = video_info['r_frame_rate'].split('/')
fps = float(fps_num) / float(fps_den)
return fps, width, height
def read_frame_time(self, video_path, frame_idx, fps, width, height=None):
frame_time = frame_idx / fps
v = ffmpeg.input(video_path, ss=frame_time)
if self.frame_sample_rate != -1:
v = ffmpeg.filter(v, filter_name='fps', fps=self.frame_sample_rate)
v, width, height = self.ffmpeg_transforms(v, width, height)
out, _ = (
v.output('pipe:', format='rawvideo', pix_fmt='rgb24', vframes=1).
run(capture_stdout=True, quiet=True, cmd=path_to_ffmpeg)
)
frame = np.frombuffer(out, np.uint8)
frame = frame.reshape([height, width, 3])
return frame
def transform(self, frames):
frames = frames / 255.
frames = torch.tensor(frames)
if self.normalize:
frames = self.normalization_transform(frames)
return frames
class MoviePy(MyLoader):
def __init__(self, **kwargs):
super().__init__(**kwargs)
if not self.random_segment_before_fps:
raise NotPossibleException('random_segment_before_fps=False does not apply here, because the times are '
'given in seconds, not frame number. The order of set_fps and subclip(s,e) does '
'not matter. First, the clip is moved to the start, then it is sampled at the '
'given fps.')
def read_video(self, video_path):
# If we do not use the context manager, remember to clip.close()
with moviepy.editor.VideoFileClip(video_path) as clip:
len_video = utils.get_duration(video_path, method='ffprobe')
frame_rate = clip.fps
ratio = self.frame_sample_rate / frame_rate if self.frame_sample_rate != -1 else 1
if self.frame_sample_rate == -1:
video_clip = clip
else:
video_clip = clip.with_fps(self.frame_sample_rate)
width = clip.w
height = clip.h
video_clip = self.moviepy_transforms(video_clip, width, height)
if self.load_format == 'random_frames':
frame_indices = self.get_frame_indices(len_video)
video = [video_clip.get_frame(idx / frame_rate) for idx in frame_indices]
video = np.stack(video)
audio = None
if self.load_audio:
raise NotPossibleException('Moviepy does not support loading audio from random frames')
elif self.load_format == 'random_segments':
segment_starts = self.get_segment_starts(len_video, ratio)
video = []
audio = [] if self.load_audio else None
for start in segment_starts:
start_time = start / frame_rate
end_time = (start + self.random_segment_duration / ratio) / frame_rate
video_seg = video_clip.subclip(start_time, end_time)
video_seg_array = np.array(list(video_seg.iter_frames()))
video.append(video_seg_array)
if self.load_audio:
audio.append(video_seg.audio.to_soundarray())
video = np.concatenate(video)
else: # load_format == 'all_video'
video = np.array(list(video_clip.iter_frames()))
audio = None
if self.load_audio:
audio = video_clip.audio.to_soundarray()
video_clip.close()
return video, audio
def moviepy_transforms(self, clip, width, height):
if self.resize:
if self.keep_aspect_ratio:
new_width, new_height = utils.get_proportional_sizes(self.short_side_size, width=width, height=height)
else:
new_width, new_height = self.short_side_size, self.short_side_size
clip = clip.resize((new_width, new_height))
width, height = new_width, new_height
if self.center_crop:
clip = clip.crop(x_center=width // 2, y_center=height // 2, width=self.crop_size, height=self.crop_size)
return clip
def transform(self, frames):
frames = frames / 255.
frames = frames.transpose(3, 0, 1, 2) # CTHW
if self.normalize:
frames = self.normalization_transform(torch.tensor(frames))
frames = torch.tensor(frames)
return frames
class PyTorchVideo(MyLoader, abc.ABC):
"""
https://github.com/facebookresearch/pytorchvideo
PytorchVideo has some other convenient classes, for example to organize the sampling of clips from a video in
https://github.com/facebookresearch/pytorchvideo/blob/main/pytorchvideo/data/clip_sampling.py
"""
decoder = None
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.fps_transform = None
def read_video(self, video_path):
if self.decoder == 'frames':
fps = utils.get_fps(video_path)
video_path_frames = video_path.replace('videos', 'frames').replace('.mp4', '')
# Optionally set multithread_io to False
video_reader = FrameVideo.from_directory(video_path_frames, fps, multithreaded_io=True)
else:
video_reader = EncodedVideo.from_path(video_path, decode_audio=self.load_audio, decoder=self.decoder)
fps = self.get_fps(video_reader, video_path)
len_video = video_reader.duration * fps
ratio = self.frame_sample_rate / fps if self.frame_sample_rate != -1 else 1
if self.load_format == 'random_frames':
frame_indices = self.get_frame_indices(len_video)
video = [video_reader.get_clip(idx / fps, (idx + 1) / fps) for idx in frame_indices]
video = torch.cat([v['video'] for v in video], axis=1)
audio = None
if self.load_audio:
raise NotPossibleException('PyTorchVideo does not support loading audio from random frames')
elif self.load_format == 'random_segments':
segment_starts = self.get_segment_starts(len_video, ratio)
video = []
audio = [] if self.load_audio else None
for start in segment_starts:
video_dict = video_reader.get_clip(start / fps, (start + self.random_segment_duration / ratio) / fps)
video_seg = video_dict['video']
if self.frame_sample_rate != -1:
video_seg = pytorchvideo.transforms.functional.uniform_temporal_subsample\
(video_seg, self.random_segment_duration)
video.append(video_seg)
if self.load_audio:
audio_segment = video_dict['audio']
audio.append(audio_segment)
video = torch.cat(video, dim=1)
else: # load_format == 'all_video'
video_dict = video_reader.get_clip(0, video_reader.duration)
video = video_dict['video']
if self.frame_sample_rate != -1:
"""
The uniform temporal subsampling is also very convenient when we have different lengths of videos and we
want to sample the same number of frames from each video. Not this exact setting.
"""
num_frames_final = int(video_reader.duration * self.frame_sample_rate)
video = pytorchvideo.transforms.functional.uniform_temporal_subsample(video, num_frames_final)
audio = video_dict['audio'] if self.load_audio else None
video_reader.close()
return video, audio
def transform(self, frames):
frames = frames / 255.
if self.resize:
if self.keep_aspect_ratio:
frames = pytorchvideo.transforms.ShortSideScale(self.short_side_size)(frames)
else:
frames = frames.permute(1, 2, 3, 0).numpy() # T, H, W, C
frames = transforms.resize_clip(frames, (self.short_side_size, self.short_side_size))
frames = torch.tensor(np.stack(frames)).permute(3, 0, 1, 2) # C, T, H, W
if self.center_crop:
# For some reason, this is applied on top of the dictionary, not the tensor
# 1 means center crop
frames = pytorchvideo.transforms.functional.uniform_crop(frames, self.crop_size, 1)
if self.normalize:
frames = pytorchvideo.transforms.Normalize(mean=mean_norm, std=std_norm)(frames)
return frames
@staticmethod
def get_fps(video_reader, video_path):
raise NotImplementedError
class PyTorchVideoFrames(PyTorchVideo):
can_load_audio = False
decoder = 'frames'
@staticmethod
def get_fps(video_reader, video_path):
return video_reader._fps
class PyTorchVideoPyAV(PyTorchVideo):
decoder = 'pyav'
@staticmethod
def get_fps(video_reader, video_path):
# Ideally there should be an attribute average_rate, which is what PyAV offers. But not implemented in
# pytorchvideo
return utils.get_fps(video_path)
class PyTorchVideoTorchvision(PyTorchVideo):
decoder = 'torchvision'
@staticmethod
def get_fps(video_reader, video_path):