-
Notifications
You must be signed in to change notification settings - Fork 64
/
dataset.py
171 lines (142 loc) · 6.57 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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
import numpy as np
import urllib.request
import os
import tarfile
import pickle
from sklearn.datasets import fetch_openml
def get_mnist():
mnist = fetch_openml('mnist_784', data_home=".")
x = mnist.data
y = mnist.target
# reshape to (#data, #channel, width, height)
x = np.reshape(x, (x.shape[0], 1, 28, 28)) / 255.
x_tr = np.asarray(x[:60000], dtype=np.float32)
y_tr = np.asarray(y[:60000], dtype=np.int32)
x_te = np.asarray(x[60000:], dtype=np.float32)
y_te = np.asarray(y[60000:], dtype=np.int32)
return (x_tr, y_tr), (x_te, y_te)
def binarize_mnist_class(y_train, y_test):
y_train_bin = np.ones(len(y_train), dtype=np.int32)
y_train_bin[y_train % 2 == 1] = -1
y_test_bin = np.ones(len(y_test), dtype=np.int32)
y_test_bin[y_test % 2 == 1] = -1
return y_train_bin, y_test_bin
def unpickle(file):
fo = open(file, 'rb')
dictionary = pickle.load(fo, encoding='latin1')
fo.close()
return dictionary
def conv_data2image(data):
return np.rollaxis(data.reshape((3, 32, 32)), 0, 3)
def get_cifar10(path="./mldata"):
if not os.path.isdir(path):
os.mkdir(path)
url = "http:https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz"
file_name = os.path.basename(url)
full_path = os.path.join(path, file_name)
folder = os.path.join(path, "cifar-10-batches-py")
# if cifar-10-batches-py folder doesn't exists, download from website
if not os.path.isdir(folder):
print("download the dataset from {} to {}".format(url, path))
urllib.request.urlretrieve(url, full_path)
with tarfile.open(full_path) as f:
def is_within_directory(directory, target):
abs_directory = os.path.abspath(directory)
abs_target = os.path.abspath(target)
prefix = os.path.commonprefix([abs_directory, abs_target])
return prefix == abs_directory
def safe_extract(tar, path=".", members=None, *, numeric_owner=False):
for member in tar.getmembers():
member_path = os.path.join(path, member.name)
if not is_within_directory(path, member_path):
raise Exception("Attempted Path Traversal in Tar File")
tar.extractall(path, members, numeric_owner=numeric_owner)
safe_extract(f, path=path)
urllib.request.urlcleanup()
x_tr = np.empty((0, 32 * 32 * 3))
y_tr = np.empty(1)
for i in range(1, 6):
fname = os.path.join(folder, "%s%d" % ("data_batch_", i))
data_dict = unpickle(fname)
if i == 1:
x_tr = data_dict['data']
y_tr = data_dict['labels']
else:
x_tr = np.vstack((x_tr, data_dict['data']))
y_tr = np.hstack((y_tr, data_dict['labels']))
data_dict = unpickle(os.path.join(folder, 'test_batch'))
x_te = data_dict['data']
y_te = np.array(data_dict['labels'])
bm = unpickle(os.path.join(folder, 'batches.meta'))
# label_names = bm['label_names']
# rehape to (#data, #channel, width, height)
x_tr = np.reshape(x_tr, (np.shape(x_tr)[0], 3, 32, 32)).astype(np.float32)
x_te = np.reshape(x_te, (np.shape(x_te)[0], 3, 32, 32)).astype(np.float32)
# normalize
x_tr /= 255.
x_te /= 255.
return (x_tr, y_tr), (x_te, y_te) # , label_names
def binarize_cifar10_class(y_train, y_test):
y_train_bin = np.ones(len(y_train), dtype=np.int32)
y_train_bin[(y_train == 2) | (y_train == 3) | (y_train == 4) | (y_train == 5) | (y_train == 6) | (y_train == 7)] = -1
y_test_bin = np.ones(len(y_test), dtype=np.int32)
y_test_bin[(y_test == 2) | (y_test == 3) | (y_test == 4) | (y_test == 5) | (y_test == 6) | (y_test == 7)] = -1
return y_train_bin, y_test_bin
def make_dataset(dataset, n_labeled, n_unlabeled):
def make_pu_dataset_from_binary_dataset(x, y, labeled=n_labeled, unlabeled=n_unlabeled):
labels = np.unique(y)
positive, negative = labels[1], labels[0]
x, y = np.asarray(x, dtype=np.float32), np.asarray(y, dtype=np.int32)
assert(len(x) == len(y))
perm = np.random.permutation(len(y))
x, y = x[perm], y[perm]
n_p = (y == positive).sum()
n_lp = labeled
n_n = (y == negative).sum()
n_u = unlabeled
if labeled + unlabeled == len(x):
n_up = n_p - n_lp
elif unlabeled == len(x):
n_up = n_p
else:
raise ValueError("Only support |P|+|U|=|X| or |U|=|X|.")
_prior = float(n_up) / float(n_u)
xlp = x[y == positive][:n_lp]
xup = np.concatenate((x[y == positive][n_lp:], xlp), axis=0)[:n_up]
xun = x[y == negative]
x = np.asarray(np.concatenate((xlp, xup, xun), axis=0), dtype=np.float32)
print(x.shape)
y = np.asarray(np.concatenate((np.ones(n_lp), -np.ones(n_u))), dtype=np.int32)
perm = np.random.permutation(len(y))
x, y = x[perm], y[perm]
return x, y, _prior
def make_pn_dataset_from_binary_dataset(x, y):
labels = np.unique(y)
positive, negative = labels[1], labels[0]
X, Y = np.asarray(x, dtype=np.float32), np.asarray(y, dtype=np.int32)
n_p = (Y == positive).sum()
n_n = (Y == negative).sum()
Xp = X[Y == positive][:n_p]
Xn = X[Y == negative][:n_n]
X = np.asarray(np.concatenate((Xp, Xn)), dtype=np.float32)
Y = np.asarray(np.concatenate((np.ones(n_p), -np.ones(n_n))), dtype=np.int32)
perm = np.random.permutation(len(Y))
X, Y = X[perm], Y[perm]
return X, Y
(x_train, y_train), (x_test, y_test) = dataset
x_train, y_train, prior = make_pu_dataset_from_binary_dataset(x_train, y_train)
x_test, y_test = make_pn_dataset_from_binary_dataset(x_test, y_test)
print("training:{}".format(x_train.shape))
print("test:{}".format(x_test.shape))
return list(zip(x_train, y_train)), list(zip(x_test, y_test)), prior
def load_dataset(dataset_name, n_labeled, n_unlabeled):
if dataset_name == "mnist":
(x_train, y_train), (x_test, y_test) = get_mnist()
y_train, y_test = binarize_mnist_class(y_train, y_test)
elif dataset_name == "cifar10":
(x_train, y_train), (x_test, y_test) = get_cifar10()
y_train, y_test = binarize_cifar10_class(y_train, y_test)
else:
raise ValueError("dataset name {} is unknown.".format(dataset_name))
xy_train, xy_test, prior = make_dataset(((x_train, y_train), (x_test, y_test)), n_labeled, n_unlabeled)
return xy_train, xy_test, prior