-
Notifications
You must be signed in to change notification settings - Fork 3
/
classifier.py
229 lines (219 loc) · 12.3 KB
/
classifier.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
import os
import time
import pickle
import scipy.io
import scipy.misc
import numpy as np
from nets import *
def load_data(image_dir, mode='train'):
image_file = 'train.pkl' if mode == 'train' else 'test.pkl'
image_dir = os.path.join(image_dir, image_file)
print('loading data: %s ...' % image_dir)
with open(image_dir, 'rb') as f:
data = pickle.load(f)
images = data['X'] / 127.5 - 1
labels = data['y']
print('finished loading data: %s!' % image_dir)
return images, labels
def run_acc(sess, prob, images, labels, data_test, label_test, batch_size):
image_size = data_test.shape[1]
idxes = list(range(len(data_test)))
np.random.shuffle(idxes)
data_test = data_test[idxes]
label_test = label_test[idxes]
total = 0
correct = 0
for i in range(int(len(data_test)/batch_size)):
data_batch = data_test[i*batch_size:(i+1)*batch_size]
label_batch = label_test[i*batch_size:(i+1)*batch_size]
p = sess.run(prob, {images: data_batch, labels: label_batch})
pred = np.argmax(p, axis=1)
total += batch_size
correct += np.sum(pred == label_batch)
if(len(data_test) % batch_size > 0):
num_left = len(data_test) % batch_size
data_batch = np.zeros([batch_size, image_size, image_size, 3])
label_batch = np.zeros([batch_size])
data_batch[:num_left, :, :, :] = data_test[-num_left:]
label_batch[:num_left] = label_test[-num_left:]
p = sess.run(prob, {images: data_batch, labels: label_batch})
pred = np.argmax(p, axis=1)
total += num_left
correct += np.sum(pred[:num_left] == label_batch[:num_left])
acc = correct/total
return acc
def train(image_size, hot_size, mode, svhn_dir, mnist_dir, model_dir, batch_size, learning_rate,
max_step):
images = tf.placeholder(tf.float32, [None, image_size, image_size, 3])
labels = tf.placeholder(tf.int32, [None])
_, logits, prob = Encoder_hot(images, image_size, hot_size, is_training=True, name=mode)
loss = slim.losses.sparse_softmax_cross_entropy(logits, labels)
optimizer = tf.train.AdamOptimizer(learning_rate)
train_op = slim.learning.create_train_op(loss, optimizer, variables_to_train=tf.trainable_variables())
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
if(mode == 'svhn'):
data_train, label_train = load_data(svhn_dir, mode='train')
data_test, label_test = load_data(svhn_dir, mode='test')
elif(mode == 'mnist'):
data_train, label_train = load_data(mnist_dir, mode='train')
data_test, label_test = load_data(mnist_dir, mode='test')
else:
data_train = []
label_train = []
data_tmp, label_tmp = load_data(svhn_dir, mode='train')
data_train.extend(list(data_tmp))
label_train.extend(list(label_tmp))
data_tmp, label_tmp = load_data(mnist_dir, mode='train')
data_train.extend(list(data_tmp))
label_train.extend(list(label_tmp))
data_train = np.array(data_train)
label_train = np.array(label_train)
svhn_data_test, svhn_label_test = load_data(svhn_dir, mode='test')
mnist_data_test, mnist_label_test = load_data(mnist_dir, mode='test')
with tf.Session(config=config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver()
print("start training...")
step = 0
start_time = time.time()
while (step < max_step):
i = step % int(data_train.shape[0] / batch_size)
if (i == 0):
idxes = list(range(len(data_train)))
np.random.shuffle(idxes)
data_train = data_train[idxes]
label_train = label_train[idxes]
data_batch = data_train[i * batch_size:(i + 1) * batch_size]
label_batch = label_train[i * batch_size:(i + 1) * batch_size]
feed_dict = {images: data_batch, labels: label_batch}
sess.run(train_op, feed_dict)
step += 1
if(step % 100 == 0):
saver.save(sess, os.path.join(model_dir, mode+'-classifier'), global_step=step)
print('model/'+mode+'-classifier-%d saved' % step)
if(mode == 'svhn' or mode == 'mnist'):
idxes = list(range(len(data_test)))
np.random.shuffle(idxes)
data_batch = data_test[idxes[:batch_size]]
label_batch = label_test[idxes[:batch_size]]
p, l = sess.run([prob, loss], {images: data_batch, labels: label_batch})
pred = np.argmax(p, axis=1)
acc = np.sum(pred == label_batch)/batch_size
print("[%d/%d]--[loss:%.3f]--[acc on %s:%.3f]--[time used:%.3f]"
%(step, max_step, l, mode, acc, (time.time()-start_time)))
else:
idxes = list(range(len(svhn_data_test)))
np.random.shuffle(idxes)
data_batch = svhn_data_test[idxes[:batch_size]]
label_batch = svhn_label_test[idxes[:batch_size]]
p, l1 = sess.run([prob, loss], {images: data_batch, labels: label_batch})
pred = np.argmax(p, axis=1)
acc_svhn = np.sum(pred == label_batch)/batch_size
idxes = list(range(len(mnist_data_test)))
np.random.shuffle(idxes)
data_batch = mnist_data_test[idxes[:batch_size]]
label_batch = mnist_label_test[idxes[:batch_size]]
p, l2 = sess.run([prob, loss], {images: data_batch, labels: label_batch})
l = (l1+l2)/2.0
pred = np.argmax(p, axis=1)
acc_mnist = np.sum(pred == label_batch)/batch_size
print("[%d/%d]--[loss:%.3f]--[acc on svhn:%.3f]--[acc on mnist:%.3f]--[time used:%.3f]" %
(step, max_step, l, acc_svhn, acc_mnist, (time.time()-start_time)))
start_time = time.time()
def eval(image_size, hot_size, mode, svhn_dir, mnist_dir, result_dir, model_path, batch_size,
learning_rate):
images = tf.placeholder(tf.float32, [None, image_size, image_size, 3])
labels = tf.placeholder(tf.int32, [None])
_, logits, prob = Encoder_hot(images, image_size, hot_size, is_training=False, name=mode)
loss = slim.losses.sparse_softmax_cross_entropy(logits, labels)
optimizer = tf.train.AdamOptimizer(learning_rate)
train_op = slim.learning.create_train_op(loss, optimizer)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
f_log = open(mode + '_eval.txt', 'w')
root_result_dir = result_dir
with tf.Session(config=config) as sess:
print("loading model...")
saver = tf.train.Saver()
saver.restore(sess, model_path)
print("done!")
for i in range(1, 101):
step = i * 200
result_dir = os.path.join(root_result_dir, str(step))
if (mode == 'svhn' or mode == 'both'):
p2s_data = np.array(pickle.load(open(os.path.join(result_dir, 'p2s_image'), 'rb')))
p2s_label = np.array(pickle.load(open(os.path.join(result_dir, 'p2s_label'), 'rb')))
s2s_data = np.array(pickle.load(open(os.path.join(result_dir, 's2s_image'), 'rb')))
s2s_label = np.array(pickle.load(open(os.path.join(result_dir, 's2s_label'), 'rb')))
t2s_data = np.array(pickle.load(open(os.path.join(result_dir, 't2s_image'), 'rb')))
t2s_label = np.array(pickle.load(open(os.path.join(result_dir, 't2s_label'), 'rb')))
if (mode == 'mnist' or mode == 'both'):
p2t_data = np.array(pickle.load(open(os.path.join(result_dir, 'p2t_image'), 'rb')))
p2t_label = np.array(pickle.load(open(os.path.join(result_dir, 'p2t_label'), 'rb')))
t2t_data = np.array(pickle.load(open(os.path.join(result_dir, 't2t_image'), 'rb')))
t2t_label = np.array(pickle.load(open(os.path.join(result_dir, 't2t_label'), 'rb')))
s2t_data = np.array(pickle.load(open(os.path.join(result_dir, 's2t_image'), 'rb')))
s2t_label = np.array(pickle.load(open(os.path.join(result_dir, 's2t_label'), 'rb')))
svhn_data_test, svhn_label_test = load_data(svhn_dir, mode='test')
mnist_data_test, mnist_label_test = load_data(mnist_dir, mode='test')
if (mode == 'svhn' or mode == 'both'):
svhn_acc = run_acc(sess, prob, images, labels, svhn_data_test, svhn_label_test,
batch_size)
p2s_acc = run_acc(sess, prob, images, labels, p2s_data, p2s_label, batch_size)
s2s_acc = run_acc(sess, prob, images, labels, s2s_data, s2s_label, batch_size)
t2s_acc = run_acc(sess, prob, images, labels, t2s_data, t2s_label, batch_size)
f_log.write("classifier acc on svhn test data: %f\n" % svhn_acc)
f_log.write("classifier acc on p2s data: %f\n" % p2s_acc)
f_log.write("classifier acc on s2s data: %f\n" % s2s_acc)
f_log.write("classifier acc on t2s data: %f\n" % t2s_acc)
if (mode == 'mnist' or mode == 'both'):
mnist_acc = run_acc(sess, prob, images, labels, mnist_data_test, mnist_label_test,
batch_size)
p2t_acc = run_acc(sess, prob, images, labels, p2t_data, p2t_label, batch_size)
t2t_acc = run_acc(sess, prob, images, labels, t2t_data, t2t_label, batch_size)
s2t_acc = run_acc(sess, prob, images, labels, s2t_data, s2t_label, batch_size)
f_log.write("classifier acc on mnist test data: %f\n" % mnist_acc)
f_log.write("classifier acc on p2t data: %f\n" % p2t_acc)
f_log.write("classifier acc on t2t data: %f\n" % t2t_acc)
f_log.write("classifier acc on s2t data: %f\n" % s2t_acc)
f_log.close()
if __name__ == "__main__":
train(image_size=32, hot_size=10, mode='mnist',
svhn_dir='/data/hhd/svhn',
mnist_dir='/data/hhd/mnist',
model_dir='/data/hhd/CrossDomainAdversarialAutoencoder-PaperCode/svhn-mnist/classifier',
batch_size=100, learning_rate=0.013, max_step=20000)
tf.reset_default_graph()
train(image_size=32, hot_size=10, mode='svhn',
svhn_dir='/data/hhd/svhn',
mnist_dir='/data/hhd/mnist',
model_dir='/data/hhd/CrossDomainAdversarialAutoencoder-PaperCode/svhn-mnist/classifier',
batch_size=100, learning_rate=0.013, max_step=20000)
tf.reset_default_graph()
train(image_size=32, hot_size=10, mode='both',
svhn_dir='/data/hhd/svhn',
mnist_dir='/data/hhd/mnist',
model_dir='/data/hhd/CrossDomainAdversarialAutoencoder-PaperCode/svhn-mnist/classifier',
batch_size=100, learning_rate=0.013, max_step=20000)
tf.reset_default_graph()
eval(image_size=32, hot_size=10, mode='mnist',
svhn_dir='/data/hhd/svhn',
mnist_dir='/data/hhd/mnist',
result_dir='/data/hhd/CrossDomainAdversarialAutoEncoder/svhn-mnist/result/1',
model_path='/data/hhd/CrossDomainAdversarialAutoencoder-PaperCode/svhn-mnist/classifier/mnist-classifier-20000',
batch_size=100, learning_rate=0.013)
tf.reset_default_graph()
eval(image_size=32, hot_size=10, mode='svhn',
svhn_dir='/data/hhd/svhn',
mnist_dir='/data/hhd/mnist',
result_dir='/data/hhd/CrossDomainAdversarialAutoEncoder/svhn-mnist/result/1',
model_path='/data/hhd/CrossDomainAdversarialAutoencoder-PaperCode/svhn-mnist/classifier/svhn-classifier-20000',
batch_size=100, learning_rate=0.013)
tf.reset_default_graph()
eval(image_size=32, hot_size=10, mode='both',
svhn_dir='/data/hhd/svhn',
mnist_dir='/data/hhd/mnist',
result_dir='/data/hhd/CrossDomainAdversarialAutoEncoder/svhn-mnist/result/1',
model_path='/data/hhd/CrossDomainAdversarialAutoencoder-PaperCode/svhn-mnist/classifier/both-classifier-20000',
batch_size=100, learning_rate=0.013)