-
Notifications
You must be signed in to change notification settings - Fork 4
/
hmdb_dataloader.py
249 lines (211 loc) · 8.44 KB
/
hmdb_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
from __future__ import absolute_import
from __future__ import division
from __future__ import unicode_literals
from __future__ import print_function
import torch as th
from torch.utils.data import Dataset
import pickle
import torch.nn.functional as F
import numpy as np
import re
import pandas as pd
from collections import defaultdict
from torch.utils.data.dataloader import default_collate
import json
import random
def name_to_stringlist(name):
change = {'claping': ['clapping']}
if name in change:
name_vec = change[name]
else:
name_vec = name.split('_')
return name_vec
class HMDB_DataLoader(Dataset):
"""MSRVTT dataset loader."""
def __init__(
self,
data_path,
we,
we_dim=300,
max_words=30,
num_frames_multiplier=5,
training=True,
tri_modal=False,
finetune_video=False,
video_interp=False
):
"""
Args:
"""
self.data = pickle.load(open(data_path, 'rb')) # contains a list of video names
self.we = we
self.we_dim = we_dim
self.max_words = max_words
self.max_video = 30
self.num_frames_multiplier = num_frames_multiplier
self.training = training
self.tri_modal = tri_modal
self.finetune_video = finetune_video
self.max_frames = 16
self.video_interp = video_interp
names = []
for vid in self.data:
names.append(vid['class'])
self.classes = sorted(set(names))
print('# Classes', len(self.classes))
self.class_embeds = []
for name in self.classes:
word_list = name_to_stringlist(name)
caption = ' '.join(word_list)
self.class_embeds.append(self._get_caption(caption))
self.class_embeds = th.stack(self.class_embeds, 0)
print('Shape of class embeds', self.class_embeds.shape)
def __len__(self):
return len(self.data)
def custom_collate(self, batch):
return default_collate(batch)
def _zero_pad_tensor(self, tensor, size):
if len(tensor) >= size:
return tensor[:size]
else:
zero = np.zeros((size - len(tensor), self.we_dim), dtype=np.float32)
return np.concatenate((tensor, zero), axis=0)
def _tokenize_text(self, sentence):
w = re.findall(r"[\w']+", str(sentence))
return w
def _words_to_we(self, words):
words = [word for word in words if word in self.we.vocab]
if words:
we = self._zero_pad_tensor(self.we[words], self.max_words)
return th.from_numpy(we)
else:
return th.zeros(self.max_words, self.we_dim)
def _get_caption(self, idx):
"""Chooses random caption if training. Uses set caption if evaluating."""
if self.training:
captions = idx
caption = self._words_to_we(self._tokenize_text(random.choice(captions)))
return caption
else:
caption = idx
return self._words_to_we(self._tokenize_text(caption))
def __getitem__(self, idx):
data = self.data[idx]
# load 2d and 3d features (features are pooled over the time dimension)
if self.finetune_video:
feat_2d = th.from_numpy(self.data[idx]['2d']).float()
feat_3d = th.from_numpy(self.data[idx]['3d']).float()
if self.video_interp:
feat_2d = F.interpolate(feat_2d.transpose(1, 0).unsqueeze(0), self.max_frames, mode='linear',
align_corners=True).squeeze(0)
feat_3d = F.interpolate(feat_3d.transpose(1, 0).unsqueeze(0), self.max_frames, mode='linear',
align_corners=True).squeeze(0)
else:
feat2d_buffer = th.zeros(self.max_frames, feat_2d.shape[-1])
feat_2d = feat_2d[:self.max_frames]
feat2d_buffer[:len(feat_2d)] = feat_2d
feat3d_buffer = th.zeros(self.max_frames, feat_3d.shape[-1])
feat_3d = feat_3d[:self.max_frames]
feat3d_buffer[:len(feat_3d)] = feat_3d
feat_2d = feat2d_buffer.transpose(1, 0)
feat_3d = feat3d_buffer.transpose(1, 0)
feat_2d = F.normalize(feat_2d, dim=0)
feat_3d = F.normalize(feat_3d, dim=0)
video = th.cat((feat_2d, feat_3d), dim=0)
else:
feat_2d = F.normalize(th.from_numpy(self.data[idx]['2d_pooled']).float(), dim=0)
feat_3d = F.normalize(th.from_numpy(self.data[idx]['3d_pooled']).float(), dim=0)
video = th.cat((feat_2d, feat_3d))
# load audio and zero pad/truncate if necessary
audio = th.FloatTensor(th.from_numpy(np.zeros((40, 1000), dtype=np.float32)))
# choose a caption
caption = ''
name = self.data[idx]['class']
if self.tri_modal:
word_list = name_to_stringlist(name)
caption = ' '.join(word_list)
caption = self._get_caption(caption)
return {'video': video, 'text': caption, 'video_id': idx,
'audio': audio, 'nframes': 32, 'class_name': name,
'class_id': th.ones(1)*self.classes.index(name),
'has_audio': th.zeros(1),
'video_name': self.data[idx]['video'],
'training': th.ones(1)*self.data[idx]['training']}
class MSRVTT_DataLoader_label(Dataset):
"""MSRVTT dataset loader."""
def __init__(
self,
data_path,
we,
pseudo_v,
pseudo_a,
we_dim=300,
max_words=30,
num_frames_multiplier=5,
training=True,
tri_modal=False,
):
"""
Args:
"""
self.data = pickle.load(open(data_path, 'rb'))
self.we = we
self.we_dim = we_dim
self.max_words = max_words
self.max_video = 30
self.num_frames_multiplier = num_frames_multiplier
self.training = training
self.tri_modal = tri_modal
self.pseudo_v = pseudo_v
self.pseudo_a = pseudo_a
def __len__(self):
return len(self.data)
def custom_collate(self, batch):
return default_collate(batch)
def _zero_pad_tensor(self, tensor, size):
if len(tensor) >= size:
return tensor[:size]
else:
zero = np.zeros((size - len(tensor), self.we_dim), dtype=np.float32)
return np.concatenate((tensor, zero), axis=0)
def _tokenize_text(self, sentence):
w = re.findall(r"[\w']+", str(sentence))
return w
def _words_to_we(self, words):
words = [word for word in words if word in self.we.vocab]
if words:
we = self._zero_pad_tensor(self.we[words], self.max_words)
return th.from_numpy(we)
else:
return th.zeros(self.max_words, self.we_dim)
def _get_caption(self, idx):
"""Chooses random caption if training. Uses set caption if evaluating."""
if self.training:
captions = self.data[idx]['caption']
caption = self._words_to_we(self._tokenize_text(random.choice(captions)))
return caption
else:
caption = self.data[idx]['eval_caption']
return self._words_to_we(self._tokenize_text(caption))
def __getitem__(self, idx):
video_id = self.data[idx]['id']
# load 2d and 3d features (features are pooled over the time dimension)
feat_2d = F.normalize(th.from_numpy(self.data[idx]['2d_pooled']).float(), dim=0)
feat_3d = F.normalize(th.from_numpy(self.data[idx]['3d_pooled']).float(), dim=0)
video = th.cat((feat_2d, feat_3d))
# load audio and zero pad/truncate if necessary
audio = self.data[idx]['audio']
target_length = 1024 * self.num_frames_multiplier
nframes = audio.numpy().shape[1]
p = target_length - nframes
if p > 0:
audio = np.pad(audio, ((0, 0), (0, p)), 'constant', constant_values=(0, 0))
elif p < 0:
audio = audio[:, 0:p]
audio = th.FloatTensor(audio)
# choose a caption
caption = ''
if self.tri_modal:
caption = self._get_caption(idx)
return {'video': video, 'text': caption, 'video_id': self.data[idx]['id'],
'audio': audio, 'nframes': nframes, 'pseudo_v': self.pseudo_v[idx], 'pseudo_a': self.pseudo_a[idx]}