-
Notifications
You must be signed in to change notification settings - Fork 0
/
resnet_init.py
268 lines (209 loc) · 9.61 KB
/
resnet_init.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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
import tensorflow as tf
from tensorflow.python.training.moving_averages import assign_moving_average
import numpy as np
import os, pdb
import cv2
import numpy as np
import random as rn
import tensorflow as tf
import threading
import time
global n_classes, ema_gp
ema_gp = []
n_classes = 40
def activation(x,name="activation"):
return tf.nn.relu(x, name=name)
def conv2d(name, l_input, w, b, s, p):
l_input = tf.nn.conv2d(l_input, w, strides=[1,s,s,1], padding=p, name=name)
l_input = l_input+b
return l_input
def max_pool(name, l_input, k, s):
return tf.nn.max_pool(l_input, ksize=[1, k, k, 1], strides=[1, s, s, 1], padding='VALID', name=name)
def norm(l_input, lsize=4, name="lrn"):
return tf.nn.lrn(l_input, lsize, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name=name)
def batchnorm(conv, out_channels):
with tf.variable_scope('bn'):
mean, var = tf.nn.moments(conv, axes=[0,1,2])
beta = tf.Variable(tf.zeros([out_channels]), name="beta")
gamma = tf.Variable(tf.truncated_normal([out_channels], stddev=0.1), name='gamma')
batch_norm = tf.nn.batch_norm_with_global_normalization(
conv, mean, var, beta, gamma, 0.001,
scale_after_normalization=True)
return batch_norm
def no_batchnorm(Ylogits, is_test, iteration, offset, convolutional=False):
return Ylogits, tf.no_op()
def initializer(in_filters, out_filters, name):
w1 = tf.get_variable(name+"W", [5, 5, in_filters, out_filters], initializer=tf.truncated_normal_initializer())
b1 = tf.get_variable(name+"B", [out_filters], initializer=tf.truncated_normal_initializer())
return w1, b1
def residual_block(in_x, in_filters, out_filters, stride, isDownSampled, name):
global ema_gp
# first convolution layer
if isDownSampled:
in_x = tf.nn.avg_pool(in_x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
x = batchnorm(in_x, in_filters)
x = activation(x)
w1, b1 = initializer(in_filters, in_filters, name+"first_res")
x = conv2d(name+'r1', x, w1, b1, 1, "SAME")
# second convolution layer
x = batchnorm(x, in_filters)
x = activation(x)
w2, b2 = initializer(in_filters, out_filters, name+"Second_res")
x = conv2d(name+'r2', x, w2, b2, 1, "SAME")
if in_filters != out_filters:
difference = out_filters - in_filters
left_pad = difference // 2
right_pad = difference - left_pad
identity = tf.pad(in_x, [[0, 0], [0, 0], [0, 0], [left_pad, right_pad]])
return x + identity
else:
return in_x + x
def ResNet(_X):
global n_classes
w1 = tf.get_variable("FirstW", [5, 5, 3, 30], initializer=tf.truncated_normal_initializer())
b1 = tf.get_variable("FirstB", [30], initializer=tf.truncated_normal_initializer())
x = conv2d('conv1', _X, w1, b1, 4, "VALID")
filters_num = [30,50,70,90]
block_num = [10,15,20,25]
l_cnt = 1
for i in range(len(filters_num)):
for j in range(block_num[i]):
x = residual_block(x, filters_num[i], filters_num[i], 1, False, 'ResidualBlock%d_%d'%(i,j))
print('[L-%d] Build %dth residual block %d with %d channels' % (l_cnt,i, j, filters_num[i]))
l_cnt +=1
if ((j==block_num[i]-1) & (i<len(filters_num)-1)):
x = residual_block(x, filters_num[i], filters_num[i+1], 2, True, 'Residualbottom%d_%d'%( i,j))
print('[L-%d] Build %dth connection layer %d from %d to %d channels' % (l_cnt, i, j, filters_num[i], filters_num[i+1]))
l_cnt +=1
x = batchnorm(x, filters_num[-1])
x = activation(x)
wo, bo=initializer(filters_num[-1], n_classes, "FinalOutput")
x = conv2d('final', x, wo, bo, 1, "SAME")
x = tf.reduce_mean(x, [1,2])
W = tf.get_variable("FinalW", [n_classes, n_classes], initializer=tf.truncated_normal_initializer())
b = tf.get_variable("FinalB", [n_classes], initializer=tf.truncated_normal_initializer())
out = tf.matmul(x, W) + b
return out
#==========================================================================
#=============Reading data in multithreading manner========================
#==========================================================================
def read_labeled_image_list(image_list_file, training_img_dir):
"""Reads a .txt file containing pathes and labeles
Args:
image_list_file: a .txt file with one /path/to/image per line
label: optionally, if set label will be pasted after each line
Returns:
List with all filenames in file image_list_file
"""
f = open(image_list_file, 'r')
filenames = []
labels = []
for line in f:
filename, label = line[:-1].split(' ')
filename = training_img_dir+filename
filenames.append(filename)
labels.append(int(label))
#print(str(filenames)+"\n")
return filenames, labels
def read_images_from_disk(input_queue, size1=128):
"""Consumes a single filename and label as a ' '-delimited string.
Args:
filename_and_label_tensor: A scalar string tensor.
Returns:
Two tensors: the decoded image, and the string label.
"""
label = input_queue[1]
fn=input_queue[0]
file_contents = tf.read_file(input_queue[0])
#example = tf.image.decode_jpeg(file_contents, channels=3)
example = tf.image.decode_png(file_contents, channels=3, name="dataset_image") # png fo rlfw
example=tf.image.resize_images(example, [size1,size1])
return example, label, fn
def setup_inputs(sess, filenames, training_img_dir, image_size=128, crop_size=100, isTest=False, batch_size=50):
# Read each image file
image_list, label_list = read_labeled_image_list(filenames, training_img_dir)
images = tf.cast(image_list, tf.string)
print(str(images)+"\n")
labels = tf.cast(label_list, tf.int64)
# Makes an input queue
if isTest is False:
isShuffle = True
numThr = 4
else:
isShuffle = False
numThr = 1
input_queue = tf.train.slice_input_producer([images, labels], shuffle=isShuffle)
image, y,fn = read_images_from_disk(input_queue)
channels = 3
image.set_shape([None, None, channels])
# Crop and other random augmentations
if isTest is False:
image = tf.image.random_flip_left_right(image)
image = tf.image.random_saturation(image, .95, 1.05)
image = tf.image.random_brightness(image, .05)
image = tf.image.random_contrast(image, .95, 1.05)
image = tf.random_crop(image, [crop_size, crop_size, 3])
image = tf.cast(image, tf.float32)/255.0
image, y,fn = tf.train.batch([image, y, fn], batch_size=batch_size, capacity=batch_size*3, num_threads=numThr, name='labels_and_images')
tf.train.start_queue_runners(sess=sess)
return image, y, fn, len(label_list)
batch_size = 20 #+ -
display_step = 10
learning_rate = tf.placeholder(tf.float32) # Learning rate to be fed
lr = 1e-4 # Learning rate start
tst = tf.placeholder(tf.bool)
iter = tf.placeholder(tf.int32)
# Setup the tensorflow...
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
print("Preparing the training & validation data...")
train_data, train_labels, filelist1, glen1 = setup_inputs(sess, "faceMix_FF_DC/train/train.txt", "faceMix_FF_DC/train/", batch_size=batch_size)
val_data, val_labels, filelist2, tlen1 = setup_inputs(sess, "faceMix_FF_DC/test/test.txt", "faceMix_FF_DC/test/", batch_size=batch_size,isTest=True)
max_iter = glen1*100
print("Preparing the training model with learning rate = %.5f..." % (lr))
with tf.variable_scope("ResNet") as scope:
pred = ResNet(train_data)
scope.reuse_variables()
valpred = ResNet(val_data)
with tf.name_scope('Loss_and_Accuracy'):
cost = tf.losses.sparse_softmax_cross_entropy(labels=train_labels, logits=pred)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost)
correct_prediction = tf.equal(tf.argmax(pred, 1), train_labels)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
top5=tf.reduce_mean(tf.cast(tf.nn.in_top_k(pred, train_labels, 5), tf.float32))
correct_prediction2 = tf.equal(tf.argmax(valpred, 1), val_labels)
accuracy2 = tf.reduce_mean(tf.cast(correct_prediction2, tf.float32))
tf.summary.scalar('Loss', cost)
tf.summary.scalar('Training_Accuracy', accuracy)
tf.summary.scalar('Top-5_accuracy', top5)
saver = tf.train.Saver()
init = tf.global_variables_initializer()
sess.run(init)
step = 0
writer = tf.summary.FileWriter("img/", sess.graph)
summaries = tf.summary.merge_all()
print("We are going to train the ImageNet model based on ResNet!!!")
while (step * batch_size) < max_iter:
epoch1=np.floor((step*batch_size)/glen1)
if (((step*batch_size)%glen1 < batch_size) & (lr==1e-3) & (epoch1 >2)):
lr /= 10
sess.run(optimizer, feed_dict={learning_rate: lr})
if (step % display_step == 1):
# calculate the loss
loss, acc, top5acc, summaries_string = sess.run([cost, accuracy,top5, summaries])
print("Iter=%d/epoch=%d, Loss=%.6f, Training Accuracy=%.6f, Top-5 Accuracy=%.6f, lr=%f" % (step*batch_size, epoch1 ,loss, acc, top5acc, lr))
writer.add_summary(summaries_string, step)
if (step % (display_step*10) == 1):
rounds = tlen1 // batch_size
valacc=[]
for k in range(rounds):
a2 = sess.run(accuracy2)
print("%.6f,"%(a2),end="")
valacc.append(a2)
print("\nIter=%d/epoch=%d, Validation Accuracy=%.6f" % (step*batch_size, epoch1 , np.mean(valacc)))
step += 1
print("Optimization Finished!")
#save_path = saver.save(sess, "tf_resnet_model.ckpt")
#print("Model saved in file: %s" % save_path)
exit()