-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathold_code_lstm.py
208 lines (183 loc) · 8.01 KB
/
old_code_lstm.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
#-*- coding:utf-8 -*-
import tensorflow as tf
import numpy as np
import os,sys,glob
import matplotlib
import analysis
if "DISPLAY" not in os.environ:
# remove Travis CI Error
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import data
import argparse
import utils
import eval
debug_flag_lv0=False
debug_flag_lv1=True
debug_flag_lv2=False
debug_flag_test=False
if __debug__ == debug_flag_lv0:
print '###debug | train.py |'
leaf_num=30
n_train=-3
batch_size=60
root_path, names, files = os.walk('./divided_log').next()
dir_paths = map(lambda name: os.path.join(root_path, name), names)
print 'dir paths : ',dir_paths[:]
print 'length',len(dir_paths)
if debug_flag_test:
dir_paths=dir_paths[:]
else:
dir_paths = dir_paths[:5]
print dir_paths[n_train:]
test_xs , test_ys=data.merge_all_data(dir_paths[n_train:])
plt.savefig('tmp.png')
plt.show()
train_xs , train_ys=data.merge_all_data(dir_paths[:n_train])
print "########################"
"""여기 위 까지 데이터 검증 완료"""
print '##### directory paths #####'
print dir_paths[n_train:]
train_xs, train_ys, test_xs, test_ys= list(data.get_specified_leaf(leaf_num , train_xs , train_ys , test_xs , test_ys))
ep=data.get_ep_all(dir_paths[n_train:] , leaf_n=leaf_num)
""" ep 하고 ap 는 데이터를 만들때 7개의 row 을 힉습시키고 그 다음 위치를 예측하는 형태로 x_data , y_data 을 만들었다
하지만 위 ep는 ap와 ep을 비교해 graph을 그리는 데 사용할 것이기 때문에 기존의 x_Data에서 불러온 ep데이터와 달리
ap와 같은 데이터 위치를 가지고있다 .
default 셋팅으로 길이는 7 너비는 3 으로 설정하였기 때문에 이 ep 데이터는 8번째 행부터 시작하며 데이터 지정한 열의 데이터를 가지고 온다.
데이터를 일일이 눈으로 보면서 검증했다
"""
assert len(ep) == len(test_ys) , len(test_ys)
min_ , max_ =data.get_min_max(train_xs, train_ys, test_xs, test_ys)
print 'min', min_ , 'max' ,max_
normalize_factor=10000.
train_xs=train_xs/normalize_factor
test_xs=test_xs/normalize_factor
train_ys=train_ys/normalize_factor
test_ys=test_ys/normalize_factor
min_ , max_ =data.get_min_max(train_xs, train_ys, test_xs, test_ys)
print 'min', min_ , 'max' ,max_
print train_xs.max()
print train_xs.min()
print test_xs.max()
print test_xs.min()
# train_xs, train_ys, test_xs, test_ys=data.normalize(train_xs, train_ys, test_xs, test_ys)
if __debug__ == debug_flag_lv1:
print 'shape train xs', np.shape(train_xs)
print 'shape test xs', np.shape(test_xs)
print 'shape train ys', np.shape(train_ys)
print 'shape test ys', np.shape(test_ys)
n, seq_length , n_col=np.shape(train_xs)
def lstm(hidden_dim):
return tf.contrib.rnn.BasicLSTMCell(
num_units=hidden_dim, state_is_tuple=True, activation=tf.tanh)
"""
parser=argparse.ArgumentParser()
parser.add_argument('--iter')
parser.add_argument('--learning_rate')
"""
data_dim=n_col
hidden_dim=30
output_dim=1
init_lr=0.01
reduced_lr1=20000
reduced_lr2=50000
reduced_lr3=80000
if debug_flag_test:
iterations=101
else:
iterations=150000
check_point=100
n_cell=3
x_ = tf.placeholder(tf.float32, [None, seq_length , data_dim] , name='x_')
y_ = tf.placeholder(tf.float32, [None, 1] , name ='y_')
lr_=tf.placeholder(tf.float32, name='learning_rate')
# build a LSTM network
cell = lstm(hidden_dim=hidden_dim)
multi_cell=tf.contrib.rnn.MultiRNNCell([lstm(hidden_dim) for _ in range(n_cell)])
outputs, _states = tf.nn.dynamic_rnn(multi_cell, x_, dtype=tf.float32)
print 'Cell shape : ',outputs
pred = tf.contrib.layers.fully_connected(
outputs[:, -1], output_dim, activation_fn=None) # We use the last cell's output
pred=tf.identity(pred , name='pred')
print 'FC layer output shape :',pred
# cost/loss
loss = tf.reduce_sum(tf.square(pred - y_) , name='loss') # sum of the squares
tf.summary.scalar('loss', loss)
tf.summary.scalar('learning_rate', lr_)
# optimizer
optimizer = tf.train.AdamOptimizer(lr_)
train = optimizer.minimize(loss , name='train_op')
if not os.path.isdir('./graph'):
os.mkdir('./graph')
with tf.Session() as sess:
merged = tf.summary.merge_all()
saver = tf.train.Saver()
train_writer = tf.summary.FileWriter(logdir='./logs/train')
test_writer = tf.summary.FileWriter(logdir='./logs/test')
init = tf.global_variables_initializer()
sess.run(init)
# Training step
train_loss=0
best_acc=0
best_loss=0
tmp_loss=0
try:
for i in range(iterations):
utils.show_processing(i,maxiter=iterations)
if i<= reduced_lr1:
learning_rate=init_lr
elif i <= reduced_lr2:
learning_rate = 0.001
elif i <= reduced_lr3:
learning_rate = 0.0001
else:
learning_rate=0.00001
if i%check_point ==0:
#batch_xs , batch_ys = data.next_batch(train_xs , train_ys , batch_size)
#print np.shape(batch_xs) , np.shape(batch_ys)
test_predict, outputs_, test_loss , merged_summaries= sess.run([pred, outputs, loss ,merged ], feed_dict={x_: test_xs , y_ : test_ys , lr_:learning_rate})
print("[step: {}] test loss: {}".format(i, test_loss))
print("[step: {}] train loss: {}".format(i, train_loss))
test_writer.add_summary(merged_summaries , i)
utils.plot_xy(test_predict=test_predict, test_ys=test_ys , savename='./graph/dynalog_result_normalize_'+str(i)+'.png')
utils.plot_xy(test_predict=test_predict*normalize_factor, test_ys=test_ys*normalize_factor,
savename='./graph/dynalog_result_orignal_' + str(i) + '.png')
acc=analysis.get_acc_with_ep(ep= ep , true = test_ys*normalize_factor , pred = test_predict*normalize_factor , error_range_percent=30)
print("[step: {}] test acc: {}".format(i, acc))
summary=tf.Summary(value = [tf.Summary.Value(tag='accuracy %s'%'test' , simple_value =float(acc))])
train_writer.add_summary(summary=summary , global_step=i )
if best_acc < acc:
best_acc = acc
tmp_loss = test_loss
saver.save(sess=sess, save_path='./models/acc_{}_loss_{}'.format(str(best_acc)[:4] , str(tmp_loss)[:4]), global_step=i)
print 'model saved'
elif best_acc == acc:
if best_loss > test_loss:
best_loss = test_loss
saver.save(sess=sess, save_path='./models/acc_{}_loss_{}.ckpt'.format(str(best_acc)[:4], str(best_loss)[:4]),
global_step=i)
print 'model saved'
saver.save(sess=sess, save_path='./models/{}'.format(str(i),global_step=i))
_, train_loss , merged_summaries = sess.run([train, loss , merged], feed_dict={x_: train_xs, y_: train_ys, lr_:learning_rate})
train_writer.add_summary(merged_summaries, i)
# Test step
test_predict, outputs_ , test_loss = sess.run([pred, outputs,loss], feed_dict={x_: test_xs,y_ : test_ys})
loss_val = sess.run(loss, feed_dict={y_: test_ys, pred: test_predict})
print outputs_, 'outputs shape', np.shape(outputs_)
print("RMSE: {}".format(loss_val))
#print test_predict
raise KeyboardInterrupt
except KeyboardInterrupt as kbi:
print '#### result ###'
pred=test_predict*normalize_factor
test_ys=test_ys*normalize_factor
analysis.analysis_result(true = test_ys , pred = pred , error_range_percent=5)
test_ys = test_ys * (max_ - min_) + min_
utils.plot_xy(test_predict=test_predict, test_ys=test_ys, savename='./dynalog_result_last' + '.png')
print '########'
print test_ys[:10]
print test_predict[:10]
sess.close()
print 'start evaluation'
eval.eval(x=test_xs , y=test_ys , error_range_percent=5 , model_path='./models/acc_{}_loss_{}-100'.format(str(best_acc)[:4] , str(tmp_loss)[:4]))
#np.save('./test_ep.npy',test_xs[])