-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathdataloader.py
447 lines (422 loc) · 21.9 KB
/
dataloader.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
import json
import os
import random
import statistics
from argparse import Namespace
from glob import glob
import av
import ipdb
import torch
import torch.utils.data as data
import torchvision
from torch.utils.data import ConcatDataset
from torchvision.datasets.video_utils import VideoClips
from torchvision.datasets.vision import VisionDataset
from tqdm import tqdm
import py12transforms as T
from sampler import DistributedSampler, UniformClipSampler, RandomClipSampler, ConcatSampler
# normalize = T.Normalize(mean=get_mean(dataset='kinetics'),
# std=get_std())
normalize = T.Normalize(mean=[0.43216, 0.394666, 0.37645],
std=[0.22803, 0.22145, 0.216989])
unnormalize = T.Unnormalize(mean=[0.43216, 0.394666, 0.37645],
std=[0.22803, 0.22145, 0.216989])
train_transform = torchvision.transforms.Compose([
T.ToFloatTensorInZeroOne(),
T.Resize((128, 171)),
T.RandomHorizontalFlip(),
normalize,
T.RandomCrop((112, 112))
])
test_transform = torchvision.transforms.Compose([
T.ToFloatTensorInZeroOne(),
T.Resize((128, 171)),
normalize,
T.CenterCrop((112, 112))
])
def get_flow_histogram(flow):
flow_magnitude = ((flow[..., 0] ** 2 + flow[..., 1] ** 2) ** 0.5).flatten()
flow_magnitude[flow_magnitude > 99] = 99
return torch.histc(flow_magnitude, min=0, max=100) / len(flow_magnitude)
histogram_flow_transform = lambda flow: get_flow_histogram(flow)
# CODE for 0,1,2 class imbalance (should go in init of dataset)
# y_tracker = Counter()
# fast_y_tracker = Counter()
# for video_idx, vid_clips in tqdm(enumerate(self.video_clips.clips), total=len(self.video_clips.clips)):
# video_path = self.video_clips.video_paths[video_idx]
# t_unit = av.open(video_path, metadata_errors='ignore').streams[0].time_base
# t_fail = sorted(self.fails_data[os.path.splitext(os.path.basename(video_path))[0]]['t'])
# t_fail = t_fail[len(t_fail) // 2]
# for clip_idx, clip in enumerate(vid_clips):
# start_pts = clip[0].item()
# end_pts = clip[-1].item()
# t_start = float(t_unit * start_pts)
# t_end = float(t_unit * end_pts)
# label = 0
# if t_start <= t_fail <= t_end:
# label = 1
# elif t_start > t_fail:
# label = 2
# y_tracker[label] += 1
# fast_y_tracker[self.video_clips.labels[video_idx][clip_idx]] += 1
# print({k: round(100 * v / sum(y_tracker.values()), 2) for k, v in y_tracker.items()})
class KineticsAndFails(VisionDataset):
FLOW_FPS = 8
def __init__(self, fails_path, kinetics_path, frames_per_clip, step_between_clips, fps, transform=None,
extensions=('.mp4',), video_clips=None, fails_only=False, val=False, balance_fails_only=False,
get_clip_times=False, fails_video_list=None, fns_to_remove=None, load_flow=False, flow_histogram=False,
fails_flow_path=None, all_fail_videos=False, selfsup_loss=None, clip_interval_factor=None,
labeled_fails=True, debug_dataset=False, anticipate_label=0, data_proportion=1, **kwargs):
self.clip_len = frames_per_clip / fps
self.clip_step = step_between_clips / fps
self.clip_interval_factor = clip_interval_factor
self.fps = fps
self.t = transform
self.load_flow = load_flow
self.flow_histogram = flow_histogram
self.video_clips = None
self.fails_path = fails_path
self.fails_flow_path = fails_flow_path
self.selfsup_loss = selfsup_loss
self.get_clip_times = get_clip_times
self.anticipate_label = anticipate_label
data_proportion = 1 if val else data_proportion
if video_clips:
self.video_clips = video_clips
else:
assert fails_path is None or fails_video_list is None
video_list = fails_video_list or glob(os.path.join(fails_path, '**', '*.mp4'), recursive=True)
if not fails_only:
kinetics_cls = torch.load("PATH/TO/kinetics_classes.pt")
kinetics_dist = torch.load("PATH/TO/dist.pt")
s = len(video_list)
for i, n in kinetics_dist.items():
n *= s
video_list += sorted(
glob(os.path.join(kinetics_path, '**', kinetics_cls[i], '*.mp4'), recursive=True))[
:round(n)]
self.video_clips = VideoClips(video_list, frames_per_clip, step_between_clips, fps)
with open("PATH/TO/borders.json") as f:
self.fails_borders = json.load(f)
with open("PATH/TO/all_mturk_data.json") as f:
self.fails_data = json.load(f)
self.fails_only = fails_only
self.t_from_clip_idx = lambda idx: (
(step_between_clips * idx) / fps, (step_between_clips * idx + frames_per_clip) / fps)
if not balance_fails_only: # no support for recompute clips after balance calc yet
self.video_clips.compute_clips(frames_per_clip, step_between_clips, fps)
if video_clips is None and fails_only and labeled_fails:
# if True:
if not all_fail_videos:
idxs = []
for i, video_path in enumerate(self.video_clips.video_paths):
video_path = os.path.splitext(os.path.basename(video_path))[0]
if video_path in self.fails_data:
idxs.append(i)
self.video_clips = self.video_clips.subset(idxs)
# if not val and balance_fails_only: # balance dataset
# ratios = {0: 0.3764, 1: 0.0989, 2: 0.5247}
self.video_clips.labels = []
self.video_clips.compute_clips(frames_per_clip, step_between_clips, fps)
for video_idx, vid_clips in tqdm(enumerate(self.video_clips.clips), total=len(self.video_clips.clips)):
video_path = self.video_clips.video_paths[video_idx]
if all_fail_videos and os.path.splitext(os.path.basename(video_path))[0] not in self.fails_data:
self.video_clips.labels.append([-1 for _ in vid_clips])
continue
t_unit = av.open(video_path, metadata_errors='ignore').streams[0].time_base
t_fail = sorted(self.fails_data[os.path.splitext(os.path.basename(video_path))[0]]['t'])
t_fail = t_fail[len(t_fail) // 2]
if t_fail < 0 or not 0.01 <= statistics.median(
self.fails_data[os.path.splitext(os.path.basename(video_path))[0]]['rel_t']) <= 0.99 or \
self.fails_data[os.path.splitext(os.path.basename(video_path))[0]]['len'] < 3.2 or \
self.fails_data[os.path.splitext(os.path.basename(video_path))[0]]['len'] > 30:
self.video_clips.clips[video_idx] = torch.Tensor()
self.video_clips.resampling_idxs[video_idx] = torch.Tensor()
self.video_clips.labels.append([])
continue
prev_label = 0
first_one_idx = len(vid_clips)
first_two_idx = len(vid_clips)
for clip_idx, clip in enumerate(vid_clips):
start_pts = clip[0].item()
end_pts = clip[-1].item()
t_start = float(t_unit * start_pts)
t_end = float(t_unit * end_pts)
label = 0
if t_start <= t_fail <= t_end:
label = 1
elif t_start > t_fail:
label = 2
if label == 1 and prev_label == 0:
first_one_idx = clip_idx
elif label == 2 and prev_label == 1:
first_two_idx = clip_idx
break
prev_label = label
self.video_clips.labels.append(
[0 for i in range(first_one_idx)] + [1 for i in range(first_one_idx, first_two_idx)] +
[2 for i in range(first_two_idx, len(vid_clips))])
if balance_fails_only and not val:
balance_idxs = []
counts = (first_one_idx, first_two_idx - first_one_idx, len(vid_clips) - first_two_idx)
offsets = torch.LongTensor([0] + list(counts)).cumsum(0)[:-1].tolist()
ratios = (1, 0.93, 1 / 0.93)
labels = (0, 1, 2)
lbl_mode = max(labels, key=lambda i: counts[i])
for i in labels:
if i != lbl_mode and counts[i] > 0:
n_to_add = round(counts[i] * ((counts[lbl_mode] * ratios[i] / counts[i]) - 1))
tmp = list(range(offsets[i], counts[i] + offsets[i]))
random.shuffle(tmp)
tmp_bal_idxs = []
while len(tmp_bal_idxs) < n_to_add:
tmp_bal_idxs += tmp
tmp_bal_idxs = tmp_bal_idxs[:n_to_add]
balance_idxs += tmp_bal_idxs
if not balance_idxs:
continue
t = torch.cat((vid_clips, torch.stack([vid_clips[i] for i in balance_idxs])))
self.video_clips.clips[video_idx] = t
vid_resampling_idxs = self.video_clips.resampling_idxs[video_idx]
try:
t = torch.cat(
(vid_resampling_idxs, torch.stack([vid_resampling_idxs[i] for i in balance_idxs])))
self.video_clips.resampling_idxs[video_idx] = t
except IndexError:
pass
self.video_clips.labels[-1] += [self.video_clips.labels[-1][i] for i in balance_idxs]
clip_lengths = torch.as_tensor([len(v) for v in self.video_clips.clips])
self.video_clips.cumulative_sizes = clip_lengths.cumsum(0).tolist()
fns_removed = 0
if fns_to_remove and not val:
for i, video_path in enumerate(self.video_clips.video_paths):
if fns_removed > len(self.video_clips.video_paths)//4:
break
video_path = os.path.splitext(os.path.basename(video_path))[0]
if video_path in fns_to_remove:
fns_removed += 1
self.video_clips.clips[i] = torch.Tensor()
self.video_clips.resampling_idxs[i] = torch.Tensor()
self.video_clips.labels[i] = []
clip_lengths = torch.as_tensor([len(v) for v in self.video_clips.clips])
self.video_clips.cumulative_sizes = clip_lengths.cumsum(0).tolist()
if kwargs['local_rank'] <= 0:
print(f'removed videos from {fns_removed} out of {len(self.video_clips.video_paths)} files')
# if not fails_path.startswith("PATH/TO/scenes"):
for i, p in enumerate(self.video_clips.video_paths):
self.video_clips.video_paths[i] = p.replace("PATH/TO/scenes",
os.path.dirname(fails_path))
self.debug_dataset = debug_dataset
if debug_dataset:
# self.video_clips = self.video_clips.subset([0])
pass
if data_proportion < 1:
rng = random.Random()
rng.seed(23719)
lbls = self.video_clips.labels
subset_idxs = rng.sample(range(len(self.video_clips.video_paths)), int(len(self.video_clips.video_paths)*data_proportion))
self.video_clips = self.video_clips.subset(subset_idxs)
self.video_clips.labels = [lbls[i] for i in subset_idxs]
def trim_borders(self, img, fn):
l, r = self.fails_borders[os.path.splitext(os.path.basename(fn))[0]]
w = img.shape[2] # THWC
if l > 0 and r > 0:
img = img[:, :, round(w * l):round(w * r)]
return img
def __len__(self):
return self.video_clips.num_clips()
def compute_clip_times(self, video_idx, clip_idx):
video_path = self.video_clips.video_paths[video_idx]
video_path = os.path.join(self.fails_path, os.path.sep.join(video_path.rsplit(os.path.sep, 2)[-2:]))
clip_pts = self.video_clips.clips[video_idx][clip_idx]
start_pts = clip_pts[0].item()
end_pts = clip_pts[-1].item()
t_unit = av.open(video_path, metadata_errors='ignore').streams[0].time_base
t_start = float(t_unit * start_pts)
t_end = float(t_unit * end_pts)
return t_start, t_end
def __getitem__(self, idx):
if self.load_flow:
video_idx, clip_idx = self.video_clips.get_clip_location(idx)
video_path = self.video_clips.video_paths[video_idx]
video_path = os.path.join(self.fails_path, os.path.sep.join(video_path.rsplit(os.path.sep, 2)[-2:]))
label = self.video_clips.labels[video_idx][clip_idx]
flow_path = os.path.join(self.fails_flow_path,
os.path.sep.join(os.path.splitext(video_path)[0].rsplit(os.path.sep, 2)[-2:]))
t_start, t_end = self.compute_clip_times(video_idx, clip_idx)
frame_start = round(t_start * self.FLOW_FPS)
n_frames = round(self.clip_len * self.FLOW_FPS)
flow = []
for frame_i in range(frame_start, frame_start + n_frames):
frame_fn = os.path.join(flow_path, f'{frame_i:06}.flo')
try:
flow.append(torch.load(frame_fn, map_location=torch.device('cpu')).permute(1, 2, 0).data.numpy())
except:
pass
while len(flow) < n_frames:
flow += flow
flow = flow[:n_frames]
flow = torch.Tensor(flow)
flow = self.trim_borders(flow, video_path)
if self.t is not None:
flow = self.t(flow)
return flow, label, (flow_path, t_start, t_end)
else:
video_idx, clip_idx = self.video_clips.get_clip_location(idx)
if self.anticipate_label:
assert not self.selfsup_loss, 'no anticipation with self supervision'
video_path = self.video_clips.video_paths[video_idx]
label = self.video_clips.labels[video_idx][clip_idx]
idx -= round(self.anticipate_label / self.clip_step)
new_video_idx, new_clip_idx = self.video_clips.get_clip_location(idx)
video, *_ = self.video_clips.get_clip(idx)
video = self.trim_borders(video, video_path)
if self.t is not None:
video = self.t(video)
new_t_start, new_t_end = self.compute_clip_times(new_video_idx, new_clip_idx)
old_t_start, old_t_end = self.compute_clip_times(video_idx, clip_idx)
if new_video_idx != video_idx or new_t_start > old_t_start:
label = -1
return video, label, (video_path, new_t_start, new_t_end, [])
video, audio, info, video_idx = self.video_clips.get_clip(idx)
video_path = self.video_clips.video_paths[video_idx]
# print(video_path)
try:
label = self.video_clips.labels[video_idx][clip_idx]
# if self.anticipate_label:
# video_path = self.video_clips.video_paths[video_idx]
# t_fail = statistics.median(self.fails_data[os.path.splitext(os.path.basename(video_path))[0]]['t'])
# t_start, t_end = self.compute_clip_times(video_idx, clip_idx)
# t_start += self.anticipate_label
# t_end += self.anticipate_label
# label = 0
# if t_start <= t_fail <= t_end:
# label = 1
# elif t_start > t_fail:
# label = 2
except:
label = -1
if label == 0 or self.fails_only: video = self.trim_borders(video, video_path)
if self.debug_dataset:
pass
# video[:] = 0
# video[..., 0] = 255
if self.t is not None:
video = self.t(video)
t_start = t_end = -1
if self.get_clip_times:
t_start, t_end = self.compute_clip_times(video_idx, clip_idx)
other = []
if self.selfsup_loss == 'pred_middle' or self.selfsup_loss == 'sort' or self.selfsup_loss == 'ctc':
k = round(self.clip_len / self.clip_step * self.clip_interval_factor)
video_l = [video]
try:
pvideo, paudio, pinfo, pvideo_idx = self.video_clips.get_clip(idx - k)
except:
pvideo_idx = -1
try:
nvideo, naudio, ninfo, nvideo_idx = self.video_clips.get_clip(idx + k)
except:
nvideo_idx = -1
t_start, _ = self.compute_clip_times(*self.video_clips.get_clip_location(idx))
try:
p_t_start, _ = self.compute_clip_times(*self.video_clips.get_clip_location(idx - k))
except:
p_t_start = 1000000000
try:
n_t_start, _ = self.compute_clip_times(*self.video_clips.get_clip_location(idx + k))
except:
n_t_start = -1000000000
# if pvideo_idx == video_idx:
# assert p_t_start < t_start, f"{t_start} <= prev video time {p_t_start}"
# if nvideo_idx == video_idx:
# assert t_start < n_t_start, f"{t_start} >= next video time {n_t_start}"
if pvideo_idx == video_idx and p_t_start < t_start:
pvideo = self.trim_borders(pvideo, video_path)
if self.t is not None:
pvideo = self.t(pvideo)
video_l.insert(0, pvideo)
else:
video_l.insert(0, torch.full_like(video, -1))
if nvideo_idx == video_idx and t_start < n_t_start:
nvideo = self.trim_borders(nvideo, video_path)
if self.t is not None:
nvideo = self.t(nvideo)
video_l.append(nvideo)
else:
video_l.append(torch.full_like(video, -1))
video_l = torch.stack(video_l)
video = video_l
other = [nvideo_idx == video_idx and pvideo_idx == video_idx]
if self.selfsup_loss == 'fps':
other = [self.fps]
other.append(idx)
return video, label, (video_path, t_start, t_end, *other)
def get_video_loader(**kwargs):
args = Namespace(**kwargs)
args.fails_video_list = None
if args.val:
args.fails_path = os.path.join(args.fails_path, 'val')
args.kinetics_path = os.path.join(args.kinetics_path, 'val')
else:
args.fails_path = os.path.join(args.fails_path, 'train')
args.kinetics_path = os.path.join(args.kinetics_path, 'train')
if args.fails_action_split:
args.fails_path = None
args.fails_video_list = torch.load(os.path.join(args.dataset_path, 'fails_action_split.pth'))[
'val' if args.val else 'train']
DEBUG = False
datasets = []
samplers = []
for fps in args.fps_list:
clips = None
args.fps = fps
args.step_between_clips = round(args.step_between_clips_sec * fps)
cache_path = os.path.join(args.dataset_path,
'{3}{2}{1}{0}{4}_videoclips.pth'.format('val' if args.val else 'train',
f'fails_only_{"all_" if args.all_fail_videos else ""}' if args.fails_only else '',
'bal_' if (
args.balance_fails_only and not DEBUG) else '',
'actions_' if args.fails_action_split else '',
f'{args.fps}fps'))
if args.cache_dataset and os.path.exists(cache_path):
clips = torch.load(cache_path)
if args.local_rank <= 0:
print(f'Loaded dataset from {cache_path}')
fns_to_remove = None
if args.flow_histogram:
args.transform = histogram_flow_transform
if args.remove_fns == 'action_based':
fns_to_remove = torch.load("PATH/TO/fails_remove_fns.pth")['action_remove']
elif args.remove_fns == 'random':
fns_to_remove = torch.load("PATH/TO/fails_remove_fns.pth")['random_remove']
dataset = KineticsAndFails(video_clips=clips, fns_to_remove=fns_to_remove, **vars(args))
if not args.val:
print(f'Dataset contains {len(dataset)} items')
if args.cache_dataset and args.local_rank <= 0 and clips is None: # and not args.fails_only
torch.save(dataset.video_clips, cache_path)
if args.val:
sampler = UniformClipSampler(dataset.video_clips,
1000000 if args.sample_all_clips else args.clips_per_video)
else:
sampler = RandomClipSampler(dataset.video_clips, 1000000 if args.sample_all_clips else args.clips_per_video)
datasets.append(dataset)
samplers.append(sampler)
if len(args.fps_list) > 1:
dataset = ConcatDataset(datasets)
sampler = ConcatSampler(samplers)
else:
dataset = datasets[0]
sampler = samplers[0]
if args.local_rank != -1:
sampler = DistributedSampler(sampler)
return data.DataLoader(
dataset=dataset,
batch_size=args.batch_size,
num_workers=args.workers,
shuffle=False,
# collate_fn=dataset.collate_fn,
sampler=sampler,
pin_memory=True,
drop_last=False
)