-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdata_loader.py
155 lines (132 loc) · 7.02 KB
/
data_loader.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
import scipy.io as sio
import h5py
import os
from PIL import Image
from torch.utils.data.dataset import Dataset
import numpy as np
from torchvision import datasets, transforms
class NDataset(Dataset):
def __init__(self, data, labels=None, transform=None, is_path=False, root='./'):
self.data = data
self.labels = labels
self.transform = transform
self.is_path = is_path
self.root = root
def __getitem__(self, index):
return self.data[index].astype('float32') if self.transform is None else self.transform(self.data[index]).float(), self.labels[index] if self.labels is not None else -1
def __len__(self):
return len(self.labels)
def load_data(data_name, K=400, view=0):
if (type(data_name) is list) and (view == -1):
ret = [[], [], [], [], [], [], [], [], [], []]
for dn in data_name:
data = load_deep_features(dn, K=K, view=view)
ret = [r + [d] for r, d in zip(ret, data[0: -1])]
return ret + [data[-1]]
elif view == -1:
ret = [[], [], [], [], [], [], [], [], [], []]
for v in range(2):
data = load_deep_features(data_name, K=K, view=v)
ret = [r + [d] for r, d in zip(ret, data[0: -1])]
return ret + [data[-1]]
else:
data_name = data_name[view] if type(data_name) is list else data_name
dd = load_deep_features(data_name, K=K, view=view)
return [[d] for d in dd[0: -1]] + [dd[-1]]
def load_deep_features(data_name, K=400, view=0):
train_transform, test_transform = None, None
if data_name == 'xmedianet':
valid_len, MAP = 4000, 0
split = 'img' if view == 0 else 'text'
path = '../datasets/XMediaNet/xmedianet_deep_doc2vec_data.h5py'
with h5py.File(path, 'r') as h:
data, labels = h['train_%s' % (split + 's_deep' if view == 0 else split)][()].astype('float32'), h['train_%ss_labels' % split][()].reshape([-1])
valid_data, valid_labels = data[-valid_len::], labels[-valid_len::]
train_data, train_labels = data[0: -valid_len], labels[0: -valid_len]
test_data, test_labels = h['test_%s' % (split + 's_deep' if view == 0 else split)][()].astype('float32'), h['test_%ss_labels' % split][()].reshape([-1])
elif data_name == 'nus_wide':
valid_len, MAP = 5000, 0
path = '../datasets/NUS-WIDE/nus_wide_deep_doc2vec_data_42941.h5py'
split = 'img' if view == 0 else 'text'
with h5py.File(path, 'r') as h:
data, labels = h['train_%s' % (split + 's_deep' if view == 0 else split)][()].astype('float32'), h['train_%ss_labels' % split][()].reshape([-1])
valid_data, valid_labels = data[-valid_len::], labels[-valid_len::]
train_data, train_labels = data[0: -valid_len], labels[0: -valid_len]
test_data, test_labels = h['test_%s' % (split + 's_deep' if view == 0 else split)][()].astype('float32'), h['test_%ss_labels' % split][()].reshape([-1])
elif data_name == 'INRIA-Websearch':
MAP = 0
split = 'img' if view == 0 else 'txt'
data = sio.loadmat('../datasets/INRIA-Websearch/INRIA-Websearch.mat')
train_data, train_labels = data['tr_%s' % split].astype('float32'), data['tr_%s_lab' % split].reshape([-1])
valid_data, valid_labels = data['val_%s' % split].astype('float32'), data['val_%s_lab' % split].reshape([-1])
test_data, test_labels = data['te_%s' % split].astype('float32'), data['te_%s_lab' % split].reshape([-1])
elif data_name.lower() == 'mnist':
MAP = None
train_dataset = datasets.MNIST('../datasets/MNIST_SVHN/MNIST/', train=True, download=True)
data, labels = train_dataset.data.numpy(), train_dataset.train_labels.numpy()
valid_data, valid_labels = data[-10000::], labels[-10000::]
train_data, train_labels = data[0: -10000], labels[0: -10000]
test_dataset = datasets.MNIST('../datasets/MNIST_SVHN/MNIST/', train=False, download=True)
test_data, test_labels = test_dataset.data.numpy(), test_dataset.test_labels.numpy()
train_transform = transforms.Compose([
lambda x: x.unsqueeze(-1),
transforms.ToPILImage(),
# transforms.RandomResizedCrop(32),
# transforms.RandomHorizontalFlip(),
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,))
])
test_transform = transforms.Compose([
lambda x: x.unsqueeze(-1),
transforms.ToPILImage(),
# transforms.Resize(32),
# transforms.CenterCrop(32),
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,))
])
elif data_name.lower() == 'svhn':
MAP = None
train_dataset = datasets.SVHN('../datasets/MNIST_SVHN/SVHN/', split='train', download=True)
data, labels = train_dataset.data, train_dataset.labels
valid_data, valid_labels = data[-10000::], labels[-10000::]
train_data, train_labels = data[0: -10000], labels[0: -10000]
test_dataset = datasets.SVHN('../datasets/MNIST_SVHN/SVHN/', split='test', download=True)
test_data, test_labels = test_dataset.data, test_dataset.labels
# inx = np.arange(train_data.shape[0])
# np.random.shuffle(inx)
# train_labeled_data, train_labeled_labels, train_unlabeled_data, train_unlabeled_labels = train_data[inx[0: K]], train_labels[inx[0: K]], train_data[inx[K::]], train_labels[inx[K::]]
train_transform = transforms.Compose([
transforms.ToPILImage(),
# transforms.RandomResizedCrop(32),
# transforms.RandomHorizontalFlip(),
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,))
])
test_transform = transforms.Compose([
transforms.ToPILImage(),
# transforms.Resize(32),
# transforms.CenterCrop(32),
transforms.Resize(32),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,))
])
inx = np.arange(train_data.shape[0])
np.random.shuffle(inx)
classes = np.unique(train_labels)
labeled_inx, unlabeled_inx = [], []
Kc = int(K // len(classes))
for c in classes:
inx = (c == train_labels).nonzero()[0]
np.random.shuffle(inx)
labeled_inx.append(inx[0: Kc])
unlabeled_inx.append(inx[Kc::])
labeled_inx = np.concatenate(labeled_inx)
unlabeled_inx = np.concatenate(unlabeled_inx)
np.random.shuffle(labeled_inx)
np.random.shuffle(unlabeled_inx)
inx = np.concatenate([labeled_inx, unlabeled_inx])
train_labeled_data, train_labeled_labels, train_unlabeled_data, train_unlabeled_labels = train_data[inx[0: K]], train_labels[inx[0: K]], train_data[inx[K::]], train_labels[inx[K::]] * 0 - 1
return train_labeled_data, train_labeled_labels, train_unlabeled_data, train_unlabeled_labels, valid_data, valid_labels, test_data, test_labels, train_transform, test_transform, MAP