-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpsenet_train_pb.py
132 lines (109 loc) · 5.63 KB
/
psenet_train_pb.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
# -*- coding: utf-8 -*-
import tensorflow as tf
import os
import time
import logging
import numpy as np
from model_net import psenet_model as PM
from model_net import model_v1
from data_pre import data_pre,pre_tools
from tensorflow.python.framework.graph_util import convert_variables_to_constants
os.environ['CUDA_VISIBLE_DEVICES']='0'
#os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
flages=tf.flags
FLAGS = flages.FLAGS
flages.DEFINE_float("learning_rate",0.0005,"learning rate")
flages.DEFINE_integer("kernal_num", 3, "kernal number")
flages.DEFINE_boolean("model_restore",False,"wheater to restore model")
flages.DEFINE_integer("batch_size", 8, "batch size")
flages.DEFINE_integer("epochs", 800, "epochs of net")
def train_PSE():
data_path = data_pre.get_data_path()
data_num = len(data_path)
num_batch = int(data_num / FLAGS.batch_size)
tf_image = tf.placeholder(dtype=tf.float32,shape=[FLAGS.batch_size,None,None,3],name="image")
tf_gt = tf.placeholder(dtype=tf.float32,shape=[FLAGS.batch_size,None,None],name="gt_text")
tf_mask = tf.placeholder(dtype=tf.float32,shape=[FLAGS.batch_size,None,None],name="mask")
tf_kernal = tf.placeholder(dtype=tf.float32,shape=[FLAGS.batch_size,FLAGS.kernal_num - 1,None,None],name="kernal_image")
resnet = PM.ResNet(PM.BottleBlock(),FLAGS.kernal_num,True,0.5)
logites = resnet(tf_image) ## (batch, 7, size, size)
#logites,_ = model_v1.model(tf_image, FLAGS.kernal_num)
#Loss = PM.Dec_Loss(logites=logites,gt_texts=tf_gt,gt_kernels=tf_kernal,training_masks=tf_mask)
Loss = PM.Dec_Loss_1(logites=logites, gt_text=tf_gt, gt_kernals=tf_kernal, train_mask=tf_mask)
global_step = tf.Variable(0, name='global_step', trainable=False)
learning_rate = tf.train.exponential_decay(
learning_rate=FLAGS.learning_rate,
global_step=global_step,
decay_steps=num_batch * 200,
decay_rate=0.5,
staircase=True)
'''
boundaries = [200, 400]
learing_rates = [0.001,0.0005,0.00025]
learning_rate = tf.train.piecewise_constant(global_step, boundaries=boundaries, values=learing_rates)
'''
optim = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss=Loss, global_step=global_step)
#ema = tf.train.ExponentialMovingAverage(0.9999,global_step)
#optim = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=0.99).minimize(loss=Loss, global_step=global_step)
#optim = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss=Loss, global_step=global_step)
tensorboard_dir = "tensorboard"
tf.summary.scalar("loss",Loss)
merged_summary = tf.summary.merge_all()
writer = tf.summary.FileWriter(tensorboard_dir)
model_save_dir = "./checkpoints/pretrain_train/"
train_start_time = time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime(time.time()))
model_name = 'PSENet_BC-{:d}_k{:d}_{:s}.ckpt'.format(FLAGS.batch_size,FLAGS.kernal_num,str(train_start_time))
model_save_path = os.path.join(model_save_dir, model_name)
logging.basicConfig(filename='./checkpoints/' + model_name + '.log',
format='%(asctime)s - %(pathname)s - %(levelname)s: %(message)s',
level=logging.DEBUG, filemode='a', datefmt='%Y-%m-%d%I:%M:%S %p')
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
sess = tf.Session(config=sess_config)
with sess.as_default():
if FLAGS.model_restore:
print("None!")
else:
step = 0
init = tf.global_variables_initializer()
sess.run(init)
writer.add_graph(sess.graph)
print("First step is:",step)
Begain_learn_rate = FLAGS.learning_rate
train_acc = 0
start_epoch = 0
if FLAGS.model_restore:
start_epoch = int(step / num_batch)
for epoch in range(start_epoch,FLAGS.epochs):
data_loader = data_pre.Data_load_pre(data_path,FLAGS.kernal_num) ### random shuffle
for i in range(num_batch):
images, gt_texts, train_masks, kernal_images = pre_tools.batch_data(data_loader, i, FLAGS.batch_size)
step += 1
_,learn_rate,loss,train_pred,merge_summary_value = sess.run([optim,learning_rate,Loss,logites,merged_summary],
feed_dict={tf_image:images,tf_gt:gt_texts,tf_mask:train_masks,tf_kernal:kernal_images})
############# 输出 learning_rate ###############################
if Begain_learn_rate != learn_rate:
information = "############ New Learning_Rate {:6f} in step {:d} ###########".format(learn_rate,step)
logging.info(information)
print(information)
Begain_learn_rate = learn_rate
if step % 20 == 0:
score_text = pre_tools.calcult_acc(train_pred[:,0,:,:],gt_texts,train_masks,pre_tools.runningScore(2))
acc = score_text['Mean Acc']
information = '## Epoch {:d} Step_Train / Total_Batch: {:d} / {:d} train_loss= {:5f} train_acc= {:5f}'. \
format(epoch, step, num_batch, loss,acc)
print(information) ### 输出到屏幕
logging.info(information) ### 输出到log文件
#if step % (num_batch * 1) == 0: ### 每 35个epoch进行一次验证，并保存最优模型
if step % 100 == 0: ### 每 35个epoch进行一次验证，并保存最优模型
#if train_acc < acc:
print("############# Save model in Epoch {:d} Step_Train / Total_Batch {:d} / {:d} ############".\
format(epoch,step,num_batch))
#saver.save(sess, model_save_path, global_step=step)
graph_def = tf.get_default_graph().as_graph_def()
output_graph_def = convert_variables_to_constants(sess, sess.graph_def, output_node_names=['transpose'])
model_f = tf.gfile.GFile("./checkpoints/pb/pse.pb","wb")
model_f.write(output_graph_def.SerializeToString())
#train_acc = acc
if __name__=="__main__":
train_PSE()