-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathdatasets.py
220 lines (173 loc) · 9.93 KB
/
datasets.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
import os
import random
import torch
import h5py
import numpy as np
import json
import math
from tqdm import tqdm
from helpers.bbox_helper import get_loc_label, get_ctr_label
from helpers.vsumm_helper import get_keyshot_summ
class MSMODataset(object):
def __init__(self, mode='train', args=None):
self.gt = json.load(open('{}/{}/annotation/{}.json'.format(args.data_root, args.dataset, mode)))
self.id_list = list(self.gt.keys())
self.video_dict = np.load('{}/{}/feature/video_resnet50_{}.npy'.format(args.data_root, args.dataset, mode), allow_pickle=True).item()
self.text_dict = np.load('{}/{}/feature/text_roberta_{}.npy'.format(args.data_root, args.dataset, mode), allow_pickle=True).item()
if args.dataset == 'Daily_Mail':
self.video_summ_dict = np.load('{}/{}/feature/video_summ_resnet50_{}.npy'.format(args.data_root, args.dataset, mode), allow_pickle=True).item()
else:
self.video_summ_dict = {}
for id in tqdm(self.id_list):
self.video_dict[id] = torch.tensor(self.video_dict[id]).to(torch.float32)
self.text_dict[id] = torch.tensor(self.text_dict[id]).to(torch.float32)
if args.dataset == 'Daily_Mail':
self.video_summ_dict[id] = torch.tensor(self.video_summ_dict[id]).to(torch.float32)
else:
self.video_summ_dict[id] = torch.zeros(1).to(torch.float32)
self.dataset = args.dataset
def __len__(self):
return len(self.id_list)
def __getitem__(self, index):
id = self.id_list[index]
video = self.video_dict[id] # [T, 2048]
video_summ = self.video_summ_dict[id]
text = self.text_dict[id] # [N, 768]
num_frame = video.shape[0]
num_keyframe = video_summ.shape[0]
num_sentence = text.shape[0]
if self.dataset == 'Daily_Mail':
video_label = torch.tensor(self.gt[id]['video_label'], dtype=torch.long)
assert torch.sum(video_label) == num_keyframe
else:
video_label = torch.zeros(num_frame).to(torch.long)
video_label[0] = 1
text_label = torch.tensor(self.gt[id]['text_label'], dtype=torch.long)
article_sentence = self.gt[id]['article_sentence']
highlight = self.gt[id]['highlight']
mask_video = torch.ones(num_frame, dtype=torch.long)
mask_video_summ = torch.ones(num_keyframe, dtype=torch.long)
mask_text = torch.ones(num_sentence, dtype=torch.long)
video_to_text_mask = torch.zeros(1)
text_to_video_mask = torch.zeros(1)
return video, video_summ, text, mask_video, mask_video_summ, mask_text, video_label, text_label, article_sentence, highlight, video_to_text_mask, text_to_video_mask
class BLiSSDataset(object):
def __init__(self, mode='train', args=None):
self.gt = json.load(open('{}/{}/annotation/{}.json'.format(args.data_root, args.dataset, mode)))
self.clip_id_list = list(self.gt.keys())
video_feature_path = '{}/{}/feature/video_clip_{}.npy'.format(args.data_root, args.dataset, mode)
text_feature_path = '{}/{}/feature/text_roberta_{}.npy'.format(args.data_root, args.dataset, mode)
video_summ_feature_path = '{}/{}/feature/video_summ_clip_{}.npy'.format(args.data_root, args.dataset, mode)
video_feature_dict = np.load(video_feature_path, allow_pickle=True).item()
text_feature_dict = np.load(text_feature_path, allow_pickle=True).item()
video_summ_feature_dict = np.load(video_summ_feature_path, allow_pickle=True).item()
self.video_dict = {}
self.video_summ_dict = {}
self.text_dict = {}
for clip_id in tqdm(self.clip_id_list):
self.video_dict[clip_id] = torch.tensor(video_feature_dict[clip_id]).to(torch.float32)
self.text_dict[clip_id] = torch.tensor(text_feature_dict[clip_id]).to(torch.float32)
self.video_summ_dict[clip_id] = torch.tensor(video_summ_feature_dict[clip_id]).to(torch.float32)
def __len__(self):
return len(self.clip_id_list)
def __getitem__(self, index):
clip_id = self.clip_id_list[index]
video_id = self.gt[clip_id]['video_id']
video = self.video_dict[clip_id] # [T, 512]
video_summ = self.video_summ_dict[clip_id] # [N, 512]
text = self.text_dict[clip_id] # [T, 768]
video_label = torch.tensor(self.gt[clip_id]['video_label'], dtype=torch.long)
text_label = torch.tensor(self.gt[clip_id]['text_label'], dtype=torch.long)
num_frame = self.gt[clip_id]['num_frame']
num_keyframe = self.gt[clip_id]['num_keyframe']
num_sentence = self.gt[clip_id]['num_sentence']
assert torch.sum(video_label) == num_keyframe
sentence = self.gt[clip_id]['sentence']
highlight = self.gt[clip_id]['highlight']
time_index = self.gt[clip_id]['sentence_time']
video_to_text_mask = torch.zeros((num_frame, num_sentence), dtype=torch.long)
text_to_video_mask = torch.zeros((num_sentence, num_frame), dtype=torch.long)
for j in range(num_sentence):
start_frame, end_frame = time_index[j]
video_to_text_mask[start_frame: end_frame, j] = 1
text_to_video_mask[j, start_frame: end_frame] = 1
mask_video = torch.ones(num_frame, dtype=torch.long)
mask_video_summ = torch.ones(num_keyframe, dtype=torch.long)
mask_text = torch.ones(num_sentence, dtype=torch.long)
return video, video_summ, text, mask_video, mask_video_summ, mask_text, video_label, text_label, sentence, highlight, video_to_text_mask, text_to_video_mask
class VideoSummDataset(object):
def __init__(self, keys, args=None):
self.keys = keys
self.video_dict = h5py.File('{}/{}/feature/eccv16_dataset_{}_google_pool5.h5'.format(args.data_root, args.dataset, args.dataset.lower()), 'r')
text_feature_path = '{}/{}/feature/text_roberta.npy'.format(args.data_root, args.dataset)
text_feature_dict = np.load(text_feature_path, allow_pickle=True).item()
video_id_list = text_feature_dict.keys()
self.text_dict = {}
for video_id in video_id_list:
self.text_dict[video_id] = torch.from_numpy(text_feature_dict[video_id]).to(torch.float32)
def __len__(self):
return len(self.keys)
def __getitem__(self, index):
key = self.keys[index]
video_name = key.split('/')[-1]
video_file = self.video_dict[video_name]
video = torch.from_numpy(video_file['features'][...].astype(np.float32)) # [T, 1024]
text = self.text_dict[video_name] # [T, 1024]
gtscore = video_file['gtscore'][...].astype(np.float32) # [T]
change_points = video_file['change_points'][...].astype(np.int32) # [S, 2], S: number of segments, each row stores indices of a segment
n_frames = video_file['n_frames'][...].astype(np.int32) # [N], N: number of frames, N = T * 15
n_frame_per_seg = video_file['n_frame_per_seg'][...].astype(np.int32) # [S], indicates number of frames in each segment
picks = video_file['picks'][...].astype(np.int32) # [T], posotions of subsampled frames in original video
user_summary = np.zeros(0, dtype=np.float32)
if 'user_summary' in video_file:
user_summary = video_file['user_summary'][...].astype(np.float32)
gtscore -= gtscore.min()
gtscore /= gtscore.max()
keyshot_summ, gtscore_upsampled = get_keyshot_summ(gtscore, change_points, n_frames, n_frame_per_seg, picks)
target = keyshot_summ[::15]
video_cls_label = target
video_loc_label = get_loc_label(target)
video_ctr_label = get_ctr_label(target, video_loc_label)
video_cls_label = torch.from_numpy(video_cls_label)
video_loc_label = torch.from_numpy(video_loc_label)
video_ctr_label = torch.from_numpy(video_ctr_label)
num_frame = video.shape[0]
num_sentence = text.shape[0]
frame_sentence_ratio = int(math.ceil(num_frame / num_sentence))
text_cls_label = np.zeros((num_sentence), dtype=bool)
for j in range(num_sentence):
start_frame = j * frame_sentence_ratio
end_frame = min((j + 1) * frame_sentence_ratio, num_frame)
if video_cls_label[start_frame: end_frame].any():
text_cls_label[j] = True
text_loc_label = get_loc_label(text_cls_label)
text_ctr_label = get_ctr_label(text_cls_label, text_loc_label)
text_cls_label = torch.from_numpy(text_cls_label)
text_loc_label = torch.from_numpy(text_loc_label)
text_ctr_label = torch.from_numpy(text_ctr_label)
video_to_text_mask = torch.zeros((num_frame, num_sentence), dtype=torch.long)
text_to_video_mask = torch.zeros((num_sentence, num_frame), dtype=torch.long)
for j in range(num_sentence):
start_frame = j * frame_sentence_ratio
end_frame = min((j + 1) * frame_sentence_ratio, num_frame)
video_to_text_mask[start_frame: end_frame, j] = 1
text_to_video_mask[j, start_frame : end_frame] = 1
mask_video = torch.ones(num_frame, dtype=torch.long)
mask_text = torch.ones(num_sentence, dtype=torch.long)
ratio = 0.15
return video, text, mask_video, mask_text, video_cls_label, video_loc_label, video_ctr_label, text_cls_label, text_loc_label, text_ctr_label, \
user_summary, n_frames, ratio, n_frame_per_seg, picks, change_points, video_to_text_mask, text_to_video_mask
def worker_init_fn(worker_id):
"""
Re-seed each worker process to preserve reproducibility
"""
worker_seed = torch.initial_seed() % 2**32
np.random.seed(worker_seed)
random.seed(worker_seed)
return
def my_collate_fn(batch):
batched_output_list = []
for i in range(len(batch[0])):
batched_output = [item[i] for item in batch]
batched_output_list.append(batched_output)
return batched_output_list