-
Notifications
You must be signed in to change notification settings - Fork 59
/
lvu.py
456 lines (414 loc) · 17.8 KB
/
lvu.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
import os
import os
import io
import random
import numpy as np
from numpy.lib.function_base import disp
import torch
from torchvision import transforms
import warnings
from decord import VideoReader, cpu
from torch.utils.data import Dataset
from .random_erasing import RandomErasing
from .video_transforms import (
Compose, Resize, CenterCrop, Normalize,
create_random_augment, random_short_side_scale_jitter,
random_crop, random_resized_crop_with_shift, random_resized_crop,
horizontal_flip, random_short_side_scale_jitter, uniform_crop,
)
from .volume_transforms import ClipToTensor
try:
from petrel_client.client import Client
has_client = True
except ImportError:
has_client = False
class LVU(Dataset):
"""Load your own video classification dataset."""
def __init__(self, anno_path, prefix='', split=' ', mode='train', clip_len=8,
frame_sample_rate=2, crop_size=224, short_side_size=256,
new_height=256, new_width=340, keep_aspect_ratio=True,
num_segment=1, num_crop=1, test_num_segment=10, test_num_crop=3,
args=None, trimmed=60, time_stride=16):
self.anno_path = anno_path
self.prefix = prefix
self.split = split
self.mode = mode
self.clip_len = clip_len
self.frame_sample_rate = frame_sample_rate
self.crop_size = crop_size
self.short_side_size = short_side_size
self.new_height = new_height
self.new_width = new_width
self.keep_aspect_ratio = keep_aspect_ratio
self.num_segment = num_segment
self.test_num_segment = test_num_segment
self.num_crop = num_crop
self.test_num_crop = test_num_crop
self.args = args
self.aug = False
self.rand_erase = False
self.trimmed = trimmed
self.time_stride = time_stride
print(f"Use trimmed videos of {trimmed} seconds")
print(f"Time stride: {time_stride} seconds")
assert num_segment == 1
if self.mode in ['train']:
self.aug = True
if self.args.reprob > 0:
self.rand_erase = True
if VideoReader is None:
raise ImportError("Unable to import `decord` which is required to read videos.")
import pandas as pd
cleaned = pd.read_csv(self.anno_path, header=None, delimiter=self.split)
self.ori_dataset_samples = list(cleaned.values[:, 0])
self.ori_label_array = list(cleaned.values[:, 1])
self.ori_duration_array = list(cleaned.values[:, 2])
# 裁剪原始视频
self.dataset_samples = []
self.label_array = []
self.start_array = []
self.end_array = []
for idx, duration in enumerate(self.ori_duration_array):
if duration < trimmed:
# 将整个视频作为一个片段
self.dataset_samples.append(self.ori_dataset_samples[idx])
self.label_array.append(self.ori_label_array[idx])
self.start_array.append(0)
self.end_array.append(duration)
else:
starts = [i for i in range(0, int(duration), time_stride)]
for start in starts:
end = start + trimmed
# 如果计算的结束点超过视频时长,检查最后一个片段是否至少为trimmed长度的一半
if end > duration:
if duration - start >= trimmed / 2:
# 如果最后一个片段长度至少为trimmed的一半,则使用实际结束时间
end = duration
else:
# 如果最后一个片段长度小于trimmed的一半,则不使用这个片段
continue
self.dataset_samples.append(self.ori_dataset_samples[idx])
self.label_array.append(self.ori_label_array[idx])
self.start_array.append(start)
self.end_array.append(end)
self.client = None
if has_client:
self.client = Client('~/petreloss.conf')
if (mode == 'train'):
pass
elif (mode == 'validation'):
self.data_transform = Compose([
Resize(self.short_side_size, interpolation='bilinear'),
CenterCrop(size=(self.crop_size, self.crop_size)),
ClipToTensor(),
Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
elif mode == 'test':
self.data_resize = Compose([
Resize(size=(short_side_size), interpolation='bilinear')
])
self.data_transform = Compose([
ClipToTensor(),
Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
self.test_seg = []
self.test_dataset = []
self.test_label_array = []
self.test_start_array = []
self.test_end_array = []
for ck in range(self.test_num_segment):
for cp in range(self.test_num_crop):
for idx in range(len(self.label_array)):
sample_label = self.label_array[idx]
self.test_label_array.append(sample_label)
self.test_start_array.append(self.start_array[idx])
self.test_end_array.append(self.end_array[idx])
self.test_dataset.append(self.dataset_samples[idx])
self.test_seg.append((ck, cp))
def __getitem__(self, index):
if self.mode == 'train':
args = self.args
sample = self.dataset_samples[index]
start = self.start_array[index]
end = self.end_array[index]
buffer = self.loadvideo_decord(sample, start, end, chunk_nb=-1) # T H W C
if len(buffer) == 0:
while len(buffer) == 0:
warnings.warn("video {} not correctly loaded during training".format(sample))
index = np.random.randint(self.__len__())
sample = self.dataset_samples[index]
start = self.start_array[index]
end = self.end_array[index]
buffer = self.loadvideo_decord(sample, start, end, chunk_nb=-1)
if args.num_sample > 1:
frame_list = []
label_list = []
index_list = []
for _ in range(args.num_sample):
new_frames = self._aug_frame(buffer, args)
label = self.label_array[index]
frame_list.append(new_frames)
label_list.append(label)
index_list.append(index)
return frame_list, label_list, index_list, {}
else:
buffer = self._aug_frame(buffer, args)
return buffer, self.label_array[index], index, {}
elif self.mode == 'validation':
sample = self.dataset_samples[index]
start = self.start_array[index]
end = self.end_array[index]
buffer = self.loadvideo_decord(sample, start, end, chunk_nb=0)
if len(buffer) == 0:
while len(buffer) == 0:
warnings.warn("video {} not correctly loaded during validation".format(sample))
index = np.random.randint(self.__len__())
sample = self.dataset_samples[index]
start = self.start_array[index]
end = self.end_array[index]
buffer = self.loadvideo_decord(sample, start, end, chunk_nb=0)
buffer = self.data_transform(buffer)
return buffer, self.label_array[index], sample.split("/")[-1].split(".")[0]
elif self.mode == 'test':
sample = self.test_dataset[index]
start = self.test_start_array[index]
end = self.test_end_array[index]
chunk_nb, split_nb = self.test_seg[index]
buffer = self.loadvideo_decord(sample, start, end, chunk_nb=chunk_nb)
while len(buffer) == 0:
warnings.warn("video {}, temporal {}, spatial {} not found during testing".format(\
str(self.test_dataset[index]), chunk_nb, split_nb))
index = np.random.randint(self.__len__())
sample = self.test_dataset[index]
start = self.test_start_array[index]
end = self.test_end_array[index]
chunk_nb, split_nb = self.test_seg[index]
buffer = self.loadvideo_decord(sample, start, end, chunk_nb=chunk_nb)
buffer = self.data_resize(buffer)
if isinstance(buffer, list):
buffer = np.stack(buffer, 0)
if self.test_num_crop == 1:
spatial_step = 1.0 * (max(buffer.shape[1], buffer.shape[2]) - self.short_side_size) / 2
spatial_start = int(spatial_step)
else:
spatial_step = 1.0 * (max(buffer.shape[1], buffer.shape[2]) - self.short_side_size) \
/ (self.test_num_crop - 1)
spatial_start = int(split_nb * spatial_step)
if buffer.shape[1] >= buffer.shape[2]:
buffer = buffer[:, spatial_start:spatial_start + self.short_side_size, :, :]
else:
buffer = buffer[:, :, spatial_start:spatial_start + self.short_side_size, :]
buffer = self.data_transform(buffer)
return buffer, self.test_label_array[index], sample.split("/")[-1].split(".")[0], \
chunk_nb, split_nb
else:
raise NameError('mode {} unkown'.format(self.mode))
def _aug_frame(
self,
buffer,
args,
):
aug_transform = create_random_augment(
input_size=(self.crop_size, self.crop_size),
auto_augment=args.aa,
interpolation=args.train_interpolation,
)
buffer = [
transforms.ToPILImage()(frame) for frame in buffer
]
buffer = aug_transform(buffer)
buffer = [transforms.ToTensor()(img) for img in buffer]
buffer = torch.stack(buffer) # T C H W
buffer = buffer.permute(0, 2, 3, 1) # T H W C
# T H W C
buffer = tensor_normalize(
buffer, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]
)
# T H W C -> C T H W.
buffer = buffer.permute(3, 0, 1, 2)
# Perform data augmentation.
scl, asp = (
[0.08, 1.0],
[0.75, 1.3333],
)
buffer = spatial_sampling(
buffer,
spatial_idx=-1,
min_scale=256,
max_scale=320,
crop_size=self.crop_size,
random_horizontal_flip=False if args.data_set == 'SSV2' else True ,
inverse_uniform_sampling=False,
aspect_ratio=asp,
scale=scl,
motion_shift=False
)
if self.rand_erase:
erase_transform = RandomErasing(
args.reprob,
mode=args.remode,
max_count=args.recount,
num_splits=args.recount,
device="cpu",
)
buffer = buffer.permute(1, 0, 2, 3)
buffer = erase_transform(buffer)
buffer = buffer.permute(1, 0, 2, 3)
return buffer
def _get_seq_frames(self, video_size, start, end, num_frames, clip_idx=-1):
# 确保 start 和 end 在视频大小范围内
start = max(0, min(start, video_size - 1))
end = max(start, min(end, video_size - 1))
# 计算裁剪后视频的实际大小
clipped_video_size = end - start + 1
seg_size = max(0., float(clipped_video_size - 1) / num_frames)
seq = []
if clip_idx == -1:
for i in range(num_frames):
start_frame = int(np.round(seg_size * i)) + start
end_frame = int(np.round(seg_size * (i + 1))) + start
idx = min(random.randint(start_frame, end_frame), end)
seq.append(idx)
else:
num_segment = 1
if self.mode == 'test':
num_segment = self.test_num_segment
duration = seg_size / (num_segment + 1)
for i in range(num_frames):
start_frame = int(np.round(seg_size * i)) + start
frame_index = start_frame + int(duration * (clip_idx + 1))
idx = min(frame_index, end)
seq.append(idx)
return seq
def loadvideo_decord(self, sample, start, end, chunk_nb=0):
"""Load video content using Decord"""
fname = sample
fname = os.path.join(self.prefix, fname)
try:
if self.keep_aspect_ratio:
if "s3://" in fname:
video_bytes = self.client.get(fname)
vr = VideoReader(io.BytesIO(video_bytes),
num_threads=1,
ctx=cpu(0))
else:
vr = VideoReader(fname, num_threads=1, ctx=cpu(0))
else:
if "s3://" in fname:
video_bytes = self.client.get(fname)
vr = VideoReader(io.BytesIO(video_bytes),
width=self.new_width,
height=self.new_height,
num_threads=1,
ctx=cpu(0))
else:
vr = VideoReader(fname, width=self.new_width, height=self.new_height,
num_threads=1, ctx=cpu(0))
fps = vr.get_avg_fps()
all_index = self._get_seq_frames(
len(vr), int(start * fps), int(end * fps),
self.clip_len, clip_idx=chunk_nb
)
vr.seek(0)
buffer = vr.get_batch(all_index).asnumpy()
return buffer
except:
print("video cannot be loaded by decord: ", fname)
return []
def __len__(self):
if self.mode != 'test':
return len(self.dataset_samples)
else:
return len(self.test_dataset)
def spatial_sampling(
frames,
spatial_idx=-1,
min_scale=256,
max_scale=320,
crop_size=224,
random_horizontal_flip=True,
inverse_uniform_sampling=False,
aspect_ratio=None,
scale=None,
motion_shift=False,
):
"""
Perform spatial sampling on the given video frames. If spatial_idx is
-1, perform random scale, random crop, and random flip on the given
frames. If spatial_idx is 0, 1, or 2, perform spatial uniform sampling
with the given spatial_idx.
Args:
frames (tensor): frames of images sampled from the video. The
dimension is `num frames` x `height` x `width` x `channel`.
spatial_idx (int): if -1, perform random spatial sampling. If 0, 1,
or 2, perform left, center, right crop if width is larger than
height, and perform top, center, buttom crop if height is larger
than width.
min_scale (int): the minimal size of scaling.
max_scale (int): the maximal size of scaling.
crop_size (int): the size of height and width used to crop the
frames.
inverse_uniform_sampling (bool): if True, sample uniformly in
[1 / max_scale, 1 / min_scale] and take a reciprocal to get the
scale. If False, take a uniform sample from [min_scale,
max_scale].
aspect_ratio (list): Aspect ratio range for resizing.
scale (list): Scale range for resizing.
motion_shift (bool): Whether to apply motion shift for resizing.
Returns:
frames (tensor): spatially sampled frames.
"""
assert spatial_idx in [-1, 0, 1, 2]
if spatial_idx == -1:
if aspect_ratio is None and scale is None:
frames, _ = random_short_side_scale_jitter(
images=frames,
min_size=min_scale,
max_size=max_scale,
inverse_uniform_sampling=inverse_uniform_sampling,
)
frames, _ = random_crop(frames, crop_size)
else:
transform_func = (
random_resized_crop_with_shift
if motion_shift
else random_resized_crop
)
frames = transform_func(
images=frames,
target_height=crop_size,
target_width=crop_size,
scale=scale,
ratio=aspect_ratio,
)
if random_horizontal_flip:
frames, _ = horizontal_flip(0.5, frames)
else:
# The testing is deterministic and no jitter should be performed.
# min_scale, max_scale, and crop_size are expect to be the same.
assert len({min_scale, max_scale, crop_size}) == 1
frames, _ = random_short_side_scale_jitter(
frames, min_scale, max_scale
)
frames, _ = uniform_crop(frames, crop_size, spatial_idx)
return frames
def tensor_normalize(tensor, mean, std):
"""
Normalize a given tensor by subtracting the mean and dividing the std.
Args:
tensor (tensor): tensor to normalize.
mean (tensor or list): mean value to subtract.
std (tensor or list): std to divide.
"""
if tensor.dtype == torch.uint8:
tensor = tensor.float()
tensor = tensor / 255.0
if type(mean) == list:
mean = torch.tensor(mean)
if type(std) == list:
std = torch.tensor(std)
tensor = tensor - mean
tensor = tensor / std
return tensor