-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathdataset.py
158 lines (118 loc) · 5.36 KB
/
dataset.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
import torch, pickle, json
from torch.utils.data import Dataset, DataLoader
def get_son2parent(csv_path):
son2parent = dict()
with open (csv_path,'r') as fp:
lines = fp.readlines()
for line in lines:
if line[-1] == '\n':
line = line[:-1]
tmp_list = line.split(',')
for i in range(len(tmp_list) - 2): # ignore last digit.
key, value = tmp_list[i], tmp_list[i+1]
son2parent[key] = value
if '' in son2parent.keys():
del son2parent['']
return son2parent
class My_DS(Dataset):
def __init__(self, X, y, emb):
self.X = X
self.y = y
self.emb = emb
def __len__(self):
return self.X.shape[0]
def __getitem__(self, index):
x_batch = self.X[index, :]
y_batch = self.y[index]
a_batch = self.emb[y_batch,:]
return x_batch, y_batch, a_batch
def get_dataloader(Xtr, Xval, ytr, yval, emb, batch_size):
train_dataset = My_DS(X=Xtr, y=ytr, emb=emb)
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True, drop_last=True)
val_dataset = My_DS(X=Xval, y=yval, emb=emb)
val_loader = DataLoader(dataset=val_dataset, batch_size=batch_size, shuffle=True, drop_last=True)
return train_loader, val_loader
def split_act_data(feat, label, fns, anno_fn): # This only for ActivityNet.
with open(anno_fn) as json_file:
data = json.load(json_file)
anno_tr = {key:value for key,value in data['database'].items() if value['annotations']}
training_inx, validation_inx = [], []
for feat_fn in fns:
# Using file name to get the corresponding clip_id, 找到文件名及对应clip_id
key = feat_fn.split('.')[0].split('_')[-1]
key = '_'.join(feat_fn.split('.')[0].split('_')[2:])
clip_inx = feat_fn.split('.')[0].split('_')[0]
clip_inx = int(clip_inx)
# Get the video annotation, i.e. the action label.
video_anno = anno_tr[key]
if video_anno['subset'] == 'validation':
training_inx.append(False), validation_inx.append(True)
elif video_anno['subset'] == 'training':
training_inx.append(True), validation_inx.append(False)
training_inx, validation_inx = torch.tensor(training_inx), torch.tensor(validation_inx)
Xtr, Xval = feat[training_inx], feat[validation_inx]
ytr, yval = label[training_inx], label[validation_inx]
assert Xtr.dim() == 2 and ytr.dim() == 1
return Xtr, Xval, ytr, yval
def get_activitynet_dataset(feat_path = './data.pickle', anno_fn = './activity_net.v1-3.json'):
# Generate X, the feature and y, the label.
with open(feat_path,'rb') as f:
data1 = pickle.load(f)
label_set = list(set(data1['label']))
label_set.sort() # make sure that the label set is sorted for one-hot encoding.
label = [label_set.index(item) for item in data1['label']]
label = torch.tensor(label)
feat, fns = data1['feat'], data1['fn']
Xtr, Xval, ytr, yval = split_act_data(feat, label, fns, anno_fn)
return Xtr, Xval, ytr, yval, label_set
def get_emb(emb_type, emb_file, label_set):
n_cls = len(label_set)
if emb_type == 'rand':
emb = torch.rand(n_cls,300)
elif emb_type == 'wacv':
with open(emb_file, 'rb') as pickle_file:
content = pickle.load(pickle_file)
emb = content['embedding']
emb = torch.tensor(emb).cuda()
elif emb_type == 'oh':
one_hot = torch.zeros(n_cls, n_cls).long()
emb = one_hot.scatter_(dim=1, index=torch.unsqueeze(torch.arange(n_cls), dim=1), src=torch.ones(n_cls, n_cls).long())
emb = emb.float()
elif emb_type == 'glove':
data2 = torch.load(emb_file, map_location='cpu')
emb_names = data2['objects']
ext_emb = data2['embeddings'] # n_cls x dim
emb = torch.zeros(n_cls,ext_emb.shape[1])
emb = emb.cuda()
for i in range(n_cls):
pos = emb_names.index(label_set[i])
emb[i] = ext_emb[pos,:]
elif emb_type == 'hyp':
data2 = torch.load(emb_file, map_location='cpu')
emb_names = data2['objects']
ext_emb = data2['embeddings'] # 272 x dim
emb = torch.zeros(n_cls,ext_emb.shape[1])
for i in range(n_cls):
pos = emb_names.index(label_set[i])
emb[i] = ext_emb[pos,:]
elif emb_type == 'cone':
data2 = torch.load(emb_file, map_location='cpu')
if type(data2) is zip:
data2 = dict(data2)
emb_names = list(data2.keys())
ext_emb = list(data2.values()) # 271 x dim, root is discarded
ext_emb = torch.tensor(ext_emb)
emb = torch.zeros(n_cls,ext_emb.shape[1])
for i in range(n_cls):
pos = emb_names.index(label_set[i])
emb[i] = ext_emb[pos,:]
return emb
def get_kineticslike_dataset(train_pth_path, valid_pth_path):
data_train = torch.load(train_pth_path)
data_val = torch.load(valid_pth_path)
label_set = list(set(data_train['label']))
label_set.sort() # 有Sort很重要
Xtr, Xval = data_train['feat'], data_val['feat']
ytr = torch.tensor([label_set.index(item )for item in data_train['label']])
yval = torch.tensor([label_set.index(item )for item in data_val['label']])
return Xtr, Xval, ytr, yval, label_set