/
dataset_kinetics.py
executable file
·223 lines (185 loc) · 7.86 KB
/
dataset_kinetics.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
"""The functions for building the Kinetics400 dataset class for pretraining
Code partially borrowed from
https://github.com/YihengZhang-CV/SeCo-Sequence-Contrastive-Learning/blob/main/dataset/dataset_kinetics_v2.py.
MIT License
Copyright (c) 2020 YihengZhang-CV
"""
import torch.utils.data
import os
import copy
import random
import torch
import numpy as np
from PIL import Image
from .dataset_builder import DATASETS
def set_rng(seed):
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
@DATASETS.register_module()
class KineticsClipFolderDataset(torch.utils.data.Dataset):
def __init__(self, root, split='train', **kwargs):
super(KineticsClipFolderDataset, self).__init__()
if '##' in root: # super resource
data_root_split = root.split('##')
assert len(data_root_split) == 2
root = data_root_split[0]
self.dataset_frame_root_ssd = os.path.join(data_root_split[1], 'data')
assert '#' not in self.dataset_frame_root_ssd
assert os.path.exists(self.dataset_frame_root_ssd)
else:
self.dataset_frame_root_ssd = None
# dataset root
if '#' in root: # multiple data resources
self.dataset_root = root.split('#')
else:
self.dataset_root = [root]
for p in self.dataset_root:
if not os.path.exists(p):
print(p)
assert False
self.dataset_root_num = len(self.dataset_root)
print('using {} data sources'.format(self.dataset_root_num))
# data frame root
self.dataset_frame_root = [os.path.join(p, split) for p in self.dataset_root]
for p in self.dataset_frame_root:
assert os.path.exists(p)
# data list file
assert split in ('train', 'val')
self.dataset_list_file = os.path.join(self.dataset_root[0], split + '.txt')
assert os.path.exists(self.dataset_list_file)
# load vid samples
self.samples = self._load_list(self.dataset_list_file)
self.transform = None
def _get_aug_frame(self, frame_root, frame_idx):
frame = Image.open(os.path.join(frame_root, 'frame_{:05d}.jpg'.format(frame_idx)))
frame.convert('RGB')
if self.transform is not None:
frame_aug = self.transform(frame)
else:
frame_aug = frame
return frame_aug
def _load_list(self, list_root):
with open(list_root, 'r') as f:
lines = f.readlines()
vids = []
for k, l in enumerate(lines):
lsp = l.strip().split(' ')
# path, frame, label
if self.dataset_frame_root_ssd is not None and os.path.exists(
os.path.join(self.dataset_frame_root_ssd, lsp[0])):
vid_root = os.path.join(self.dataset_frame_root_ssd, lsp[0])
else:
vid_root = os.path.join(self.dataset_frame_root[k % self.dataset_root_num], lsp[0])
vid_root, _ = os.path.splitext(vid_root)
# use splitetxt twice because there are some video root like: abseiling/9EnSwbXxu5g.mp4.webm
vid_root, _ = os.path.splitext(vid_root)
vids.append((vid_root, int(lsp[1]), int(lsp[2])))
return vids
def __len__(self):
return len(self.samples)
def __getitem__(self, item):
raise NotImplementedError
@DATASETS.register_module()
class KineticsClipFolderDatasetMultiFrames(KineticsClipFolderDataset):
def __init__(self, root, transform=None, split='train', sample_num=0):
super(KineticsClipFolderDatasetMultiFrames, self).__init__(root, split)
self.transform = transform
self.sample_num = sample_num
assert self.transform is not None
def __getitem__(self, item):
frame_root, frame_num, cls = self.samples[item]
sample_num = frame_num if self.sample_num <= 0 or self.sample_num > frame_num else self.sample_num
frame_indices = np.round(np.linspace(1, frame_num, num=sample_num)).astype(np.int64)
frames = torch.cat([self._get_aug_frame(frame_root, frame_indices[i]) for i in range(sample_num)], dim=0)
return frames, cls
@DATASETS.register_module()
class KineticsClipFolderDatasetOrderTSN(KineticsClipFolderDataset):
def __init__(self, root, transform=None, split='train_list'):
super(KineticsClipFolderDatasetOrderTSN, self).__init__(root, split)
self.transform = transform
assert self.transform is not None
self.num_segments = 3
def __getitem__(self, item):
frame_root, frame_num, cls = self.samples[item]
initial_seed = random.randint(0, 2 ** 31)
set_rng(initial_seed)
###### Step 1: TSN samples ######
# segments (base on num_images_to_return)
frame_indices = np.round(np.linspace(1, frame_num, num=frame_num)).astype(np.int64)
segments_length = frame_num // self.num_segments
segments = []
for i in range(self.num_segments):
start_idx = i * segments_length
if i == self.num_segments - 1:
segment = frame_indices[start_idx:]
else:
end = (i + 1) * segments_length
segment = frame_indices[start_idx:end]
segments.append(segment)
# sample frames from each segments
key_images = []
queue_images = []
# debug
key_ids = []
queue_ids = []
for segment in segments:
image_path_inds = np.random.choice(segment, 2, replace=False)
for ii, ind in enumerate(image_path_inds):
image = self._get_aug_frame(frame_root, ind).unsqueeze(dim=0)
if ii == 0:
key_images.append(image)
key_ids.append(ind)
else:
queue_images.append(image)
queue_ids.append(ind)
if len(key_images) < self.num_segments:
return None
###### Step 2: SeCo samples ######
rand_segment = random.randint(0, 1)
if rand_segment == 0:
frame1_aug1 = queue_images[0].squeeze(dim=0)
frame1_aug2 = self._get_aug_frame(frame_root, queue_ids[0])
frame2_aug = queue_images[1].squeeze(dim=0)
frame3_aug = queue_images[2].squeeze(dim=0)
else:
frame1_aug1 = queue_images[2].squeeze(dim=0)
frame1_aug2 = self._get_aug_frame(frame_root, queue_ids[2])
frame2_aug = queue_images[0].squeeze(dim=0)
frame3_aug = queue_images[1].squeeze(dim=0)
###### Step 3: Order samples ######
# 4 labels: 0 (00), 1 (10), 2 (01), 3 (11)
rand_shuffle1 = random.randint(0, 1)
rand_shuffle2 = random.randint(0, 1)
if rand_shuffle1:
queue_images, queue_ids = shuffle_list(queue_images, queue_ids)
if rand_shuffle2:
key_images, key_ids = shuffle_list(key_images, key_ids)
order_label = 3 # label: 11
else:
order_label = 1 # label: 10
else:
if rand_shuffle2:
key_images, key_ids = shuffle_list(key_images, key_ids)
order_label = 2 # label: 01
else:
order_label = 0 # label: 00
# tsn q and k
tsn_q = torch.cat(queue_images, dim=0)
tsn_k = torch.cat(key_images, dim=0)
return frame1_aug1, frame1_aug2, frame2_aug, frame3_aug, order_label, tsn_q, tsn_k
def shuffle_list(l, l_idx):
l_idx_forward = copy.copy(l_idx)
l_idx_backward = copy.copy(l_idx)
l_idx_backward.reverse()
i = 0
while True:
seed = random.randint(0, 2 ** 31)
set_rng(seed)
random.shuffle(l)
set_rng(seed)
random.shuffle(l_idx)
# after shuffling, still keep the order
if l_idx != l_idx_forward and l_idx != l_idx_backward:
break
return l, l_idx