-
Notifications
You must be signed in to change notification settings - Fork 2
/
nets.py
251 lines (193 loc) · 8.07 KB
/
nets.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
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
from __future__ import division
from __future__ import print_function
import time
import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras.layers.core import Dropout
from layers import FCVarDropout, Conv2DVarDropout
from loss import sgvlb
from utils import get_minibatches_idx, clip
batch_size = 32
eps = 1e-8
class Net():
def fit(self,X,y,sess):
max_epochs = 20
# Split into training and validation sets
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.33, random_state=42)
for epoch in range(max_epochs):
start = time.time()
train_indices = get_minibatches_idx(len(X_train), batch_size, shuffle=True)
print("\nEpoch %d" % (epoch+1))
train_accs = []
for c,it in enumerate(train_indices):
batch_train_x = [X_train[i] for i in it]
batch_train_y = [y_train[i] for i in it]
feed_dict = {self.x: batch_train_x,
self.y: batch_train_y,
self.deterministic: False}
_,acc = sess.run([self.train_step,self.accuracy], feed_dict)
train_accs.append(acc)
#print(c,len(train_indices),acc)
print("Training accuracy: %.3f" % np.mean(train_accs))
val_pred = self.predict(X_val,sess)
y = np.argmax(y_val,axis=1)
val_acc = np.mean(np.equal(val_pred,y))
print("Val accuracy: %.3f" % val_acc)
print("Time taken: %.3fs" % (time.time() - start))
return
def predict(self,X,sess):
indices = get_minibatches_idx(len(X), batch_size, shuffle=False)
pred = []
for i in indices:
batch_x = [X[j] for j in i]
feed_dict = {self.x: batch_x,
self.deterministic: True}
pred_batch = sess.run(self.pred, feed_dict)
pred.append(pred_batch)
pred = np.concatenate(pred,axis=0)
pred = np.argmax(pred,axis=1)
pred = np.reshape(pred,(-1))
return pred
class LeNet(Net):
def __init__(self, img_size, num_channels, num_classes):
self.x = tf.placeholder(tf.float32, [None,img_size,img_size,num_channels], 'x')
self.y = tf.placeholder(tf.float32, [None,num_classes], 'y')
self.deterministic = tf.placeholder(tf.bool, name='d')
h = Conv2D(32, kernel_size=(3,3),
activation='relu',
input_shape=[None,img_size,img_size,num_channels])(self.x)
h = Conv2D(64, (3, 3), activation='relu')(h)
h = MaxPooling2D(pool_size=(2,2))(h)
h = Flatten()(h)
h = Dense(500, activation='relu')(h)
self.pred = Dense(num_classes, activation='softmax')(h)
pred = tf.clip_by_value(self.pred,eps,1-eps)
loss = -tf.reduce_sum(tf.log(pred)*self.y)
correct_prediction = tf.equal(tf.argmax(self.y, 1), tf.argmax(self.pred, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')
optimizer = tf.train.AdamOptimizer()
self.train_step = optimizer.minimize(loss)
class LeNetVarDropout(Net):
def __init__(self, img_size, num_channels, num_classes):
self.x = tf.placeholder(tf.float32, [None,img_size,img_size,num_channels], 'x')
self.y = tf.placeholder(tf.float32, [None,num_classes], 'y')
self.deterministic = tf.placeholder(tf.bool, name='d')
d = self.deterministic
h = Conv2DVarDropout(num_channels, 32, (3,3), strides=(1,1))(self.x,d)
h = Conv2DVarDropout(32, 64, (3,3), strides=(1,1))(h,d)
h = MaxPooling2D(pool_size=(2,2))(h)
h = Flatten()(h)
if num_channels == 1:
#h = FCVarDropout(9216,500)(h,d)
h = Dense(500)(h)
elif num_channels == 3:
h = FCVarDropout(12544,500)(h,d)
else:
raise NotImplementedError
#self.pred = FCVarDropout(500,num_classes,tf.nn.softmax)(h,d)
self.pred = Dense(num_classes,activation='softmax')(h)
pred = tf.clip_by_value(self.pred,eps,1-eps)
W = tf.get_collection('W')
log_sigma2 = tf.get_collection('log_sigma2')
loss = sgvlb(pred, self.y, W, log_sigma2, batch_size, rw=1)
correct_prediction = tf.equal(tf.argmax(self.y, 1), tf.argmax(self.pred, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')
optimizer = tf.train.AdamOptimizer()
self.train_step = optimizer.minimize(loss)
class VGG(Net):
def __init__(self, img_size, num_channels, num_classes, dropout_prob=0.0):
# Based on https://github.com/fchollet/keras/blob/master/keras/applications/vgg16.py
self.x = tf.placeholder(tf.float32, [None,img_size,img_size,num_channels], 'x')
self.y = tf.placeholder(tf.float32, [None,num_classes], 'y')
self.deterministic = tf.placeholder(tf.bool, name='d')
d = self.deterministic
phase = tf.logical_not(d)
def conv_bn(h, num_filters, phase):
h = Conv2D(num_filters, (3,3), padding='same')(h) # Linear
h = tf.contrib.layers.batch_norm(h, center=True, scale=False, is_training=phase)
return tf.nn.relu(h)
# Block 1
h = conv_bn(self.x,64,phase)
h = conv_bn(h,64,phase)
h = MaxPooling2D((2, 2), strides=(2,2))(h)
# Block 2
h = conv_bn(h,128,phase)
h = conv_bn(h,128,phase)
h = MaxPooling2D((2, 2), strides=(2,2))(h)
# Block 3
h = conv_bn(h,256,phase)
h = conv_bn(h,256,phase)
h = conv_bn(h,256,phase)
h = MaxPooling2D((2,2), strides=(2,2))(h)
# Block 4
h = conv_bn(h,512,phase)
h = conv_bn(h,512,phase)
h = conv_bn(h,512,phase)
h = MaxPooling2D((2,2), strides=(2,2))(h)
# Block 5
h = conv_bn(h,512,phase)
h = conv_bn(h,512,phase)
h = conv_bn(h,512,phase)
h = MaxPooling2D((2,2), strides=(2,2))(h)
h = Flatten()(h)
self.pred = Dense(num_classes, activation='softmax')(h)
pred = tf.clip_by_value(self.pred,eps,1-eps)
loss = -tf.reduce_sum(tf.log(pred)*self.y)
correct_prediction = tf.equal(tf.argmax(self.y, 1), tf.argmax(self.pred, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')
optimizer = tf.train.AdamOptimizer(0.001)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# Ensures that we execute the update_ops before performing the train_step
self.train_step = optimizer.minimize(loss)
class VGGVarDropout(Net):
def __init__(self, img_size, num_channels, num_classes):
# Based on https://github.com/fchollet/keras/blob/master/keras/applications/vgg16.py
self.x = tf.placeholder(tf.float32, [None,img_size,img_size,num_channels], 'x')
self.y = tf.placeholder(tf.float32, [None,num_classes], 'y')
self.deterministic = tf.placeholder(tf.bool, name='d')
d = self.deterministic
phase = tf.logical_not(d)
def conv_bn(h, filters_in, filters_out, d, phase):
h = Conv2DVarDropout(filters_in, filters_out, (3,3), padding='SAME', nonlinearity=tf.identity)(h,d) # Linear
h = tf.contrib.layers.batch_norm(h, center=True, scale=False, is_training=phase)
return tf.nn.relu(h)
# Block 1
h = conv_bn(self.x, num_channels, 64, d, phase)
h = conv_bn(h, 64, 64, d, phase)
h = MaxPooling2D((2, 2), strides=(2,2))(h)
# Block 2
h = conv_bn(h, 64, 128, d, phase)
h = conv_bn(h, 128, 128, d, phase)
h = MaxPooling2D((2, 2), strides=(2,2))(h)
# Block 3
h = conv_bn(h, 128, 256, d, phase)
h = conv_bn(h, 256, 256, d, phase)
h = conv_bn(h, 256, 256, d, phase)
h = MaxPooling2D((2,2), strides=(2,2))(h)
# Block 4
h = conv_bn(h, 256, 512, d, phase)
h = conv_bn(h, 512, 512, d, phase)
h = conv_bn(h, 512, 512, d, phase)
h = MaxPooling2D((2, 2), strides=(2, 2))(h)
# Block 5
h = conv_bn(h, 512, 512, d, phase)
h = conv_bn(h, 512, 512, d, phase)
h = conv_bn(h, 512, 512, d, phase)
h = MaxPooling2D((2, 2), strides=(2, 2))(h)
h = Flatten()(h)
self.pred = FCVarDropout(512, num_classes, tf.nn.softmax)(h,d)
pred = tf.clip_by_value(self.pred,eps,1-eps)
W = tf.get_collection('W')
log_sigma2 = tf.get_collection('log_sigma2')
loss = sgvlb(pred, self.y, W, log_sigma2, batch_size)
correct_prediction = tf.equal(tf.argmax(self.y, 1), tf.argmax(self.pred, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')
optimizer = tf.train.AdamOptimizer(0.0001)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# Ensures that we execute the update_ops before performing the train_step
self.train_step = optimizer.minimize(loss)