-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathtrain_char_seg.py
154 lines (113 loc) · 8.23 KB
/
train_char_seg.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
import os
import math
import tensorflow as tf
import cv2
import numpy as np
import random
from data_generator import get_batch
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
class Train_char_seg(object):
def __init__(self):
self.input_shape = (2048,64, 3)
self.learningrate=0.0001
self.epochs=20
self.batch_size=16
self.save_steps=100
self.one_epoch_num=10000
self.alpha=0.9
self.beta=0.1
def network(self):
network = {}
network["inputs"] = tf.placeholder(tf.float32, [self.batch_size, self.input_shape[1],self.input_shape[0], self.input_shape[2]],name='inputs')
network["down-conv1"] = tf.layers.conv2d(inputs=network["inputs"], filters=32, kernel_size=(2, 2), padding="same",activation=tf.nn.relu, name="down-conv1")
network["down-pool1"] = tf.layers.max_pooling2d(inputs=network["down-conv1"], pool_size=[2, 2], strides=2)
network["down-conv2"] = tf.layers.conv2d(inputs=network["down-pool1"], filters=64, kernel_size=(2, 2), padding="same",activation=tf.nn.relu, name="down-conv2")
network["down-pool2"] = tf.layers.max_pooling2d(inputs=network["down-conv2"], pool_size=[2, 2], strides=2)
network["down-conv3"] = tf.layers.conv2d(inputs=network["down-pool2"], filters=128, kernel_size=(2, 2), padding="same",activation=tf.nn.relu, name="down-conv3")
network["down-pool3"] = tf.layers.max_pooling2d(inputs=network["down-conv3"], pool_size=[2, 2], strides=2)
network["down-conv4"] = tf.layers.conv2d(inputs=network["down-pool3"], filters=256, kernel_size=(2, 2), padding="same",activation=tf.nn.relu, name="down-conv4")
network["down-pool4"] = tf.layers.max_pooling2d(inputs=network["down-conv4"], pool_size=[2, 2], strides=2)
network["down-conv5"] = tf.layers.conv2d(inputs=network["down-pool4"], filters=512, kernel_size=(2, 2), padding="same",activation=tf.nn.relu, name="down-conv5")
network["down-pool5"] = tf.layers.max_pooling2d(inputs=network["down-conv5"], pool_size=[2, 2], strides=2)
network["down-conv6"] = tf.layers.conv2d(inputs=network["down-pool5"], filters=512, kernel_size=(2, 2), padding="same",activation=tf.nn.relu, name="down-conv6")
network["down-pool6"] = tf.layers.max_pooling2d(inputs=network["down-conv6"], pool_size=[2, 2], strides=2)
network["up-conv1"] = tf.layers.conv2d_transpose(inputs=network["down-pool6"], filters=512, kernel_size=(1, 2),strides=(1, 2), padding="valid",activation=tf.nn.relu, name="up-conv1")
network["up-conv2"] = tf.layers.conv2d_transpose(inputs=network["up-conv1"], filters=512, kernel_size=(1, 2),strides=(1, 2), padding="valid",activation=tf.nn.relu, name="up-conv2")
network["up-conv3"] = tf.layers.conv2d_transpose(inputs=network["up-conv2"], filters=256, kernel_size=(1, 2),strides=(1, 2), padding="valid",activation=tf.nn.relu, name="up-conv3")
network["up-conv4"] = tf.layers.conv2d_transpose(inputs=network["up-conv3"], filters=128, kernel_size=(1, 2),strides=(1, 2), padding="valid",activation=tf.nn.relu, name="up-conv4")
network["up-conv5"] = tf.layers.conv2d_transpose(inputs=network["up-conv4"], filters=64, kernel_size=(1, 2), strides=(1, 2),padding="valid",activation=tf.nn.relu, name="up-conv5")
network["up-conv6"] = tf.layers.conv2d_transpose(inputs=network["up-conv5"], filters=1, kernel_size=(1, 2), strides=(1, 2),padding="valid",activation=None, name="up-conv6")
network["outputs"] =tf.contrib.layers.flatten(network["up-conv6"])
return network
def Dynamic_Weighted_Binary_CrossEntropy_loss(self,y_true, y_pred, alpha, beta):
"""
Heuristic Rules for Dynamic Loss
L(p; q) = −α Xi;qi=1log pi − β Xi;qi=0log(1 − pi)
accpos = Xi;qi=11(pi > 0:5)= Xi;qi=11;
accneg = Xi;qi=01(pi < 0:5)= Xi;qi=01:
"""
y_pred=tf.nn.sigmoid(y_pred)
y_pred=tf.cast(y_pred>0.5,tf.float32)
acc_pos_fenzi = tf.reduce_sum(tf.multiply(y_pred, y_true))
acc_pos_fenmu = tf.reduce_sum(y_true)
acc_neg_fenzi = tf.reduce_sum(tf.multiply((1.0 - y_pred), (1.0 - y_true)))
acc_neg_fenmu = tf.reduce_sum(1.0 - y_true)
acc_pos = tf.div(acc_pos_fenzi ,tf.add(acc_pos_fenmu , 1e-10))
acc_neg = tf.div(acc_neg_fenzi ,tf.add(acc_neg_fenmu , 1e-10))
seigema = tf.minimum(beta, 0.001)
alpha_new=tf.where(tf.less(acc_pos ,acc_neg),tf.add(alpha,seigema),tf.subtract(alpha , seigema))
alpha_op=tf.assign(alpha,alpha_new)
beta_new=tf.where(tf.less(acc_pos ,acc_neg),tf.subtract(beta ,seigema),tf.add(beta , seigema))
beta_op=tf.assign(beta,beta_new)
accuracy = tf.div(tf.add(acc_pos_fenzi , acc_neg_fenzi), tf.add(acc_pos_fenmu , tf.add(acc_neg_fenmu , 1e-10)))
return alpha, beta, accuracy,alpha_op,beta_op
def train(self):
#network
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(learning_rate=self.learningrate,
global_step=global_step,
decay_steps=2000,
decay_rate=0.1,
staircase=True)
model=self.network()
alpha =tf.Variable(self.alpha, trainable=False)
beta =tf.Variable(self.beta, trainable=False)
labels = tf.placeholder(tf.float32, [self.batch_size, self.input_shape[0]], name='labels')
loss = -tf.reduce_mean(alpha * (labels) * tf.log(tf.sigmoid(model["outputs"]) + 1e-10) + beta * (1.0 - labels) * tf.log(1.0 - tf.sigmoid(model["outputs"]) + 1e-10))
#loss = tf.reduce_mean(tf.losses.mean_squared_error(labels=labels,predictions=model["outputs"]))
alpha, beta,accuracy,alpha_op,beta_op=self.Dynamic_Weighted_Binary_CrossEntropy_loss(labels, model["outputs"], alpha, beta)
grad_update = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss,global_step=global_step)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
#tensorboard
tf.summary.scalar("loss", loss)
tf.summary.scalar("accuracy", accuracy)
tf.summary.scalar("alpha", alpha)
tf.summary.scalar("beta", beta)
merge_summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as session:
session.run(init)
saver = tf.train.Saver(tf.global_variables(), max_to_keep=20)
#saver.restore(session, "./save/char_seg.ckpt-800")
#tensorboard
summary_writer = tf.summary.FileWriter("./summary/", session.graph)
epoch=0
data_generator=get_batch(num_workers=16,image_width=self.input_shape[0],image_height=self.input_shape[1],image_channel=self.input_shape[2], batch_size=self.batch_size)
while True:
data = next(data_generator)
x_batch = data[0]
y_batch = data[1]
feed = {model["inputs"]: x_batch,labels: y_batch}
#print(x_batch.shape,y_batch.shape)
learning_rate_train,loss_train,alpha_train, beta_train,accuracy_train,step,summary,_,_,_=session.run([learning_rate,loss,alpha, beta,accuracy,global_step,merge_summary,grad_update,alpha_op,beta_op], feed_dict=feed)
print("learning rate:%f epoch:%d iter:%d loss:%f alpha_train:%f beta_train:%f accuracy:%f"%(learning_rate_train,epoch,step,loss_train,alpha_train, beta_train,accuracy_train))
#tensorboard
summary_writer.add_summary(summary, step)
if step > 0 and step % self.save_steps == 0:
save_path = saver.save(session, "save/char_seg.ckpt", global_step=step)
print(save_path)
if step > 0:
epoch=step*self.batch_size//self.one_epoch_num
if __name__=="__main__":
char_seg=Train_char_seg()
char_seg.train()