-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathpdNet.py
330 lines (286 loc) · 14.3 KB
/
pdNet.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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
# -*- coding: utf-8 -*-
"""
Created on Mon Mar 13 08:52:21 2017
@author: root
"""
#from __feature__ import division
import numpy as np
import tensorflow as tf
from PIL import Image
#import math
#from TFreader import *
import time
import os
IMG_H = 128
IMG_W = 64
IMG_CH = 3
LABEL_W = 2
NUM_EPOCHS = 10
BATCH_SIZE = 50
DECAY_STEP = 4000
SEED = 66478 # Set to None for random seed.
checkpoint_path = os.path.join('./', 'model.ckpt')
def data_type():
return tf.float32
class pdNet:
# Set to None for random seed.
trainFileName = None
testFileName = None
batchSize = None
numEpochs = None
graph = None
session = None
coord = None
threads = None
merge = None
train = True
def __init__(self,train_file_name,test_file_name,batch_size,num_epochs=None):
self.trainFileName = train_file_name
self.testFileName = test_file_name
self.batchSize = batch_size
self.numEpochs = num_epochs
# self.graph = tf.Graph()
# self.coord = tf.train.Coordinator()
# self.defineGraph()
# self.session = tf.Session(graph=self.graph)
# self.writer = tf.summary.FileWriter('./board', self.graph)
# self.merge = tf.summary.merge_all()
print('TFreader is initialized.')
# def __del__(self):
def read_and_decode(self,filename,num_epochs=None):
filename_queue = tf.train.string_input_producer([filename],
num_epochs=num_epochs)
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(serialized_example,
features={
'label': tf.FixedLenFeature([2], tf.int64),
'img_raw' : tf.FixedLenFeature([], tf.string),
})
img = tf.decode_raw(features['img_raw'], tf.uint8)
img = tf.reshape(img, [IMG_H, IMG_W, 3])
img = tf.cast(img, tf.float32) * (1. / 255) - 0.5
index=tf.cast(features['label'], tf.int32)
# print label[1]
return img, index
def defineGraph(self):
# with self.graph.as_default():
if self.train :
img,labels = self.read_and_decode(self.trainFileName,num_epochs=self.numEpochs)
else:
img,labels = self.read_and_decode(self.testFileName,num_epochs=self.numEpochs)
self.imgBatch,self.labelsBatch = tf.train.shuffle_batch([img,labels],
# num_epochs = self.numEpochs,
batch_size = self.batchSize,
capacity = self.batchSize*3,
min_after_dequeue = 100)
#---- net variables
# eval_data = tf.place
with tf.name_scope('model'):
#This is where training samples and labels are fed to the graph.
#These placeholder nodes will be fed a batch of training data at each
#training step using the (feed_dict) argument to the Run() call below.
self.train_data_node = tf.placeholder(
data_type(),
shape=(BATCH_SIZE,IMG_H,IMG_W,IMG_CH),
name = 'train_data_node')
self.train_labels_node = tf.placeholder(data_type(),shape=(BATCH_SIZE,LABEL_W),
name = 'train_labels_node')
c1_w = tf.Variable(
tf.truncated_normal([8,8,IMG_CH,64],#5x5 filter depth 32.
stddev=0.1,
seed=SEED,dtype=data_type(),
name = 'c1_w'))
c1_b = tf.Variable(tf.zeros([64],dtype = data_type(),
name = 'c1_b'))
c2_w = tf.Variable(
tf.truncated_normal([4,4,64,32],
stddev=0.1,
seed=SEED,dtype = data_type(),
name = 'c2_w'))
c2_b = tf.Variable(tf.zeros([32],dtype = data_type(),
name = 'c2_b'))
fc1_w = tf.Variable(
tf.truncated_normal([IMG_H//4*IMG_W//4*32,1024],
stddev=0.1,
seed=SEED,dtype = data_type(),
name = 'fc1_w'))
fc1_b = tf.Variable(tf.constant(0.1,shape=[1024],dtype=data_type(),
name = 'fc1_b'))
fc2_w = tf.Variable(
tf.truncated_normal([1024,LABEL_W],
stddev=0.1,
seed=SEED,dtype = data_type(),
name = 'fc2_w'))
fc2_b = tf.Variable(tf.constant(0.1,shape=[LABEL_W],dtype=data_type(),
name = 'fc2_b'))
#-----------------------------
tf.summary.histogram('c1_w',c1_w)
#--------------------- end net variables
#
def model(data,train=False):
# shape matches the data layout:[image index,y,x,depth].
c1 = tf.nn.conv2d(data,
c1_w,
strides=[1,1,1,1],
padding = 'SAME')
# Bias and rectified linear non-linearity.
relu1 = tf.nn.relu(tf.nn.bias_add(c1,c1_b))
# Max pooling
pool1 = tf.nn.max_pool(relu1,
ksize=[1,2,2,1],
strides=[1,2,2,1],
padding='SAME')
c2 = tf.nn.conv2d(pool1,
c2_w,
strides=[1,1,1,1],
padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(c2,c2_b))
pool2 = tf.nn.max_pool(relu2,
ksize=[1,2,2,1],
strides=[1,2,2,1],
padding='SAME')
# Reshape the feature map cuboid into a 2D matrix to feed it to the fully connected layers.
poolShape = pool2.get_shape().as_list()
reshape = tf.reshape(
pool2,
[poolShape[0],poolShape[1]*poolShape[2]*poolShape[3]])
# Fully connected layer. Note that the '+' operation automatically broadcasts the biases.
fc1 = tf.nn.tanh(tf.matmul(reshape,fc1_w) + fc1_b)
# Add a 50% dropout during training training only.
# Dropout also scales activations such that no rescaling is needed at evaluation time
if train:
fc1 = tf.nn.dropout(fc1,0.5,seed=SEED)
# return tf.nn.sigmoid(tf.matmul(fc1,fc2_w) + fc2_b)
return (tf.matmul(fc1,fc2_w) + fc2_b)
#--------------------end model
# Training computation: logits + cross-entropy loss
logits = model(self.train_data_node,self.train)
self.loss1 = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=logits,labels=self.train_labels_node))
# L2 regularization for the fully connected parameters.
regularizers = (tf.nn.l2_loss(fc1_w) + tf.nn.l2_loss(fc1_b) + tf.nn.l2_loss(fc2_w) + tf.nn.l2_loss(fc2_b))
# Add the regularization term to the loss.
#
# if tf.is_nan(regularizers) is not None:
# self.loss = self.loss1*(1 + tf.nn.tanh(1e-2*regularizers))
# else:
# self.loss = self.loss1
# self.loss = self.loss1 + 2e-7*regularizers
## regularizers = tf.nn.tanh(regularizers)
## self.loss = self.loss1 + regularizers
self.loss = self.loss1
tf.summary.histogram('loss1',self.loss1)
tf.summary.histogram('loss',self.loss)
# Optimizer: set up a variable that's incremented once per batch and controls the learning rate decay.
batch = tf.Variable(0,dtype=data_type())
# Decay once per epoch, using an exponential schedule starting at 0.01
self.learningRate = tf.train.exponential_decay(
0.05, # Base learning rate.
batch * self.batchSize, # Current index into the dataset
DECAY_STEP, # Decay step.
0.99, # Decay rate.
staircase=True)
# learningRate = 0.1
self.optimizer = tf.train.MomentumOptimizer(self.learningRate,0.9).minimize(self.loss,global_step=batch)
# Predictions for the current training minibatch
self.trainPrediction = tf.nn.sigmoid(logits)
# self.trainPrediction = (logits)
# self.merge = tf.merge_v2_checkpoints()
# self.merge = tf.merge_all_summaries()
#------------ END def defineGraph()
#
def error_rate(self,predictions, labels):
"""Return the error rate based on dense predictions and sparse labels."""
# num = len(predictions)
# pmax = np.argmax(predictions, 1)
# lmax = np.argmax(labels, 1)
# s = np.sum(pmax==lmax)
# rate = 1 - s/num
# return rate
return 100.0 - (
100.0 *
np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1)) /
np.float32(len(predictions)))
def train(self):
with tf.Graph().as_default() as graph:
self.train = True
self.defineGraph()
saver = tf.train.Saver(tf.all_variables())
session = tf.Session(graph=graph)
with session :
tf.local_variables_initializer().run(session=session) # epoch计数变量是local variable
tf.global_variables_initializer().run(session=session)
coord = tf.train.Coordinator()
# threads = tf.train.start_queue_runners(sess=session, coord=coord)
tf.train.start_queue_runners(sess=session)
try:
step = 0
while not coord.should_stop():
startTime = time.time()
# with session.as_default():
imgBatch_r,labelBatch_r = session.run([self.imgBatch,self.labelsBatch])#session.run([tfReader.imgBatch,tfReader.labelsBatch])
feed_dict = {self.train_data_node: imgBatch_r,
self.train_labels_node: labelBatch_r}
_,lrate,l1,l,prediction = session.run([self.optimizer,
# self.merge,
self.learningRate,
self.loss1,
self.loss,
self.trainPrediction],
feed_dict = feed_dict)
elapsed_time = time.time() - startTime
err = self.error_rate(prediction,labelBatch_r)
if not step%10:
print('step:%d time:%f s' %(step,elapsed_time))
# print('time:%f s' %elapsed_time)
print('loss1:%f ,loss:%f ,learnrate:%f' %(l1,l,lrate))
print('error_rate: %f' %err)
step += 1
except tf.errors.OutOfRangeError:
print('Train is over.')
finally:
print('test')
coord.request_stop()
# coord.join(threads)
# self.session.close()
saver.save(session, checkpoint_path)
print("Trained and saved.")
def evalu(self):
with tf.Graph().as_default() as graph:
self.train = False
self.defineGraph()
saver = tf.train.Saver()
session = tf.Session(graph=graph)
with session.as_default():
saver.restore(session, checkpoint_path)
# tf.local_variables_initializer().run(session=self.session) # epoch计数变量是local variable
# tf.global_variables_initializer().run(session=self.session)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=session, coord=coord)
try:
predict = []
goldLabel = []
while not coord.should_stop():
# startTime = time.time()
# with self.session.as_default():
imgBatch_r,labelBatch_r = session.run([self.imgBatch,self.labelsBatch])#session.run([tfReader.imgBatch,tfReader.labelsBatch])
feed_dict = {self.train_data_node: imgBatch_r,
self.train_labels_node: labelBatch_r}
re = session.run([self.trainPrediction],
feed_dict = feed_dict)
predict.append(re)
goldLabel.append(labelBatch_r)
except tf.errors.OutOfRangeError:
print('Evaluation is over.')
finally:
coord.request_stop()
coord.join(threads)
err = self.error_rate(predict,goldLabel)
print('evalu_err_rate: %f' %err)
session.close()
if __name__ == '__main__':
net = pdNet("trainData_64x128.tfrecords","testData_64x128.tfrecords",BATCH_SIZE,num_epochs=NUM_EPOCHS)
# net.train()
net.evalu()
del net