-
Notifications
You must be signed in to change notification settings - Fork 2
/
datahandler.py
210 lines (167 loc) · 9.25 KB
/
datahandler.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
"""
This module contains the class DataHandler, which handles NUS-WIDE and COCO datasets
for multilabel classification.
"""
import os
import copy
import random
random.seed(26)
import h5py
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
import params
class DataHandler(object):
"""
The class keeps handles for the data and labels, and builds generators to feed them.
"""
no_classes = {'nus_wide': 81,
'ms_coco': 80}
data_types = {'nus_wide': ('train', 'test'),
'ms_coco': ('train', 'val')}
preprocessor_aug = ImageDataGenerator(featurewise_center=True,
horizontal_flip=True,
width_shift_range=1.0/8,
height_shift_range=1.0/8,
zoom_range=1.0/8,
fill_mode='constant')
preprocessor = ImageDataGenerator(featurewise_center=True)
def __init__(self, dataset, labeled_ratio, corruption_ratio):
self.dataset = dataset
self.labeled_ratio = labeled_ratio
self.corruption_ratio = corruption_ratio
f_dataset = h5py.File(os.path.join('dataset', dataset + '.h5'), 'r')
self.train_images = f_dataset[self.data_types[dataset][0] + '_images']
self.train_image_shapes = f_dataset[self.data_types[dataset][0] + '_image_shapes']
self.train_labels = f_dataset[self.data_types[dataset][0] + '_labels']
self.val_images = f_dataset[self.data_types[dataset][1] + '_images']
self.val_image_shapes = f_dataset[self.data_types[dataset][1] + '_image_shapes']
self.val_labels = f_dataset[self.data_types[dataset][1] + '_labels']
self.mixed_labels = [] # to be fed from outside
self.preprocessor_aug.mean = f_dataset['mean']
self.preprocessor.mean = f_dataset['mean']
# decide on which examples are labeled
if self.labeled_ratio == 100:
self.inds_labeled = range(len(self.train_labels))
self.inds_unlabeled = []
else:
approx_labeled_ratio = self.labeled_ratio
while True:
self.inds_labeled = []
no_examples_per_class = np.sum(self.train_labels, axis=0)
no_labeled_examples_per_class = (np.round(no_examples_per_class * (approx_labeled_ratio / 100.0)).astype(np.int))
for ind_class in range(self.train_labels.shape[1]):
inds = np.where(self.train_labels[:, ind_class] == 1)[0].tolist()
random.shuffle(inds)
self.inds_labeled += inds[:no_labeled_examples_per_class[ind_class]]
self.inds_labeled = list(set(self.inds_labeled))
true_labeled_ratio = (float(len(self.inds_labeled)) / len(self.train_labels)) * 100
# have 0.5% tolerance, err on the positive side
if true_labeled_ratio > self.labeled_ratio + 5:
approx_labeled_ratio -= 2.5
elif true_labeled_ratio > self.labeled_ratio + 0.5:
approx_labeled_ratio -= 0.25
elif true_labeled_ratio < self.labeled_ratio:
approx_labeled_ratio += 0.25
else:
break
self.inds_unlabeled = [x for x in range(len(self.train_labels)) if x not in self.inds_labeled]
self.inds_labeled_sorted = list(self.inds_labeled)
self.inds_labeled_sorted.sort()
# corrupt labels
if self.corruption_ratio > 0:
corrupted_labels = self.train_labels[:]
inds_corrupted = copy.deepcopy(self.inds_labeled)
random.shuffle(inds_corrupted)
inds_corrupted = inds_corrupted[:int(len(inds_corrupted) * (self.corruption_ratio / 100.0))]
class_prob = np.sum(self.train_labels, axis=0) / self.train_labels.shape[0]
for ind_labeled in inds_corrupted:
pos_classes = np.zeros(len(class_prob), dtype=np.bool)
while np.sum(pos_classes) == 0:
pos_prob = 1 - np.random.rand(len(class_prob))
pos_classes = class_prob > pos_prob
corrupted_labels[ind_labeled] = 0
corrupted_labels[ind_labeled, pos_classes] = 1
self.train_labels = corrupted_labels
def generator(self, data_type, aug, shuffle_batches=True):
"""
A generator to be used for training/prediction/etc. in keras.
"""
if data_type == 'train_mixed':
while True:
no_labeled = len(self.inds_labeled)
no_unlabeled = len(self.inds_unlabeled)
no_examples = no_labeled + no_unlabeled
no_batches = int(np.ceil(float(no_examples) / params.batch_size))
inds = range(no_batches * params.batch_size)
inds[no_examples:] = range(len(inds) - no_examples)
random.shuffle(inds)
for ind_batch in range(no_batches):
inds_image = inds[ind_batch * params.batch_size:(ind_batch + 1) * params.batch_size]
images_flat = np.empty((params.batch_size, self.train_images[0].shape[0]), dtype=np.float32)
if aug:
prep = self.preprocessor_aug
else:
prep = self.preprocessor
labels_batch = np.empty((params.batch_size, self.train_labels.shape[1] + 1), dtype=np.float32)
for ind, ind_image in enumerate(inds_image):
if ind_image < no_labeled:
images_flat[ind] = self.train_images[self.inds_labeled[ind_image]]
labels_batch[ind] = self.mixed_labels[self.inds_labeled[ind_image]]
else:
ind_image_offset = ind_image - no_labeled
images_flat[ind] = self.train_images[self.inds_unlabeled[ind_image_offset]]
labels_batch[ind] = self.mixed_labels[self.inds_unlabeled[ind_image_offset]]
images_batch = np.empty((params.batch_size, 224, 224, 3), dtype=np.float32)
for ind_image, image_flat in enumerate(images_flat):
shaped_image = image_flat.reshape(224, 224, 3)
images_batch[ind_image] = prep.random_transform(shaped_image)
yield images_batch, labels_batch
else:
while True:
if data_type == 'train_all': # not used
no_examples = len(self.train_images)
elif data_type == 'train_labeled':
no_examples = len(self.inds_labeled)
elif data_type == 'train_labeled_sorted':
no_examples = len(self.inds_labeled_sorted)
elif data_type == 'train_unlabeled': # not used
no_examples = len(self.inds_unlabeled)
elif data_type == 'val':
no_examples = len(self.val_images)
no_batches = int(np.ceil(float(no_examples) / params.batch_size))
inds_batch = range(no_batches)
if shuffle_batches:
random.shuffle(inds_batch)
for ind_batch in inds_batch:
if ind_batch == no_batches - 1:
inds = range(no_examples - params.batch_size, no_examples)
else:
inds = range(ind_batch * params.batch_size, (ind_batch + 1) * params.batch_size)
if data_type == 'train_all':
images_flat = self.train_images[inds]
labels_batch = self.train_labels[inds]
elif data_type == 'train_labeled':
inds = [self.inds_labeled[i] for i in inds]
inds.sort()
images_flat = self.train_images[inds]
labels_batch = self.train_labels[inds]
elif data_type == 'train_labeled_sorted':
inds = [self.inds_labeled_sorted[i] for i in inds]
images_flat = self.train_images[inds]
labels_batch = self.train_labels[inds]
elif data_type == 'train_unlabeled':
inds = [self.inds_unlabeled[i] for i in inds]
images_flat = self.train_images[inds]
labels_batch = self.train_labels[inds]
elif data_type == 'val':
images_flat = self.val_images[inds]
labels_batch = self.val_labels[inds]
if aug:
prep = self.preprocessor_aug
else:
prep = self.preprocessor
images_batch = np.empty((params.batch_size, 224, 224, 3), dtype=np.float32)
for ind_image, image_flat in enumerate(images_flat):
shaped_image = image_flat.reshape(224, 224, 3)
images_batch[ind_image] = prep.random_transform(shaped_image)
yield images_batch, labels_batch