質問編集履歴
2
大幅に質問内容を変更しました
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -1,329 +1,4 @@
|
|
|
1
1
|
###前提・実現したいこと
|
|
2
2
|
Kerasのライブラリを使い，GRUに複数の時系列データを学習させたいのですが手段がわかりません．
|
|
3
3
|
現在は，時系列データを１パターンだけ学習させ，テストデータを与えるところまでは出ています．
|
|
4
|
-
具体的には，時系列データのdatファイルが30個あり，15個を学習データ，15個をテストデータとして用いたいのです．
|
|
4
|
+
具体的には，時系列データのdatファイルが30個あり，15個を学習データ，15個をテストデータとして用いたいのです．
|
|
5
|
-
|
|
6
|
-
|
|
7
|
-
###該当のソースコード
|
|
8
|
-
```python
|
|
9
|
-
import numpy as np
|
|
10
|
-
import matplotlib.pyplot as plt
|
|
11
|
-
from keras.models import Sequential
|
|
12
|
-
from keras.layers.core import Dense, Activation
|
|
13
|
-
from keras.layers.recurrent import GRU
|
|
14
|
-
from keras.optimizers import Adam
|
|
15
|
-
from keras.callbacks import EarlyStopping
|
|
16
|
-
from sklearn.model_selection import train_test_split
|
|
17
|
-
import os
|
|
18
|
-
|
|
19
|
-
np.random.seed(0)
|
|
20
|
-
|
|
21
|
-
def zscore(x, axis = None):
|
|
22
|
-
xmean = x.mean(axis=axis, keepdims=True)
|
|
23
|
-
xstd = np.std(x, axis=axis, keepdims=True)
|
|
24
|
-
zscore = (x-xmean)/xstd
|
|
25
|
-
return zscore
|
|
26
|
-
|
|
27
|
-
def read(N=10, T=200):
|
|
28
|
-
a=np.loadtxt('pre10.dat',delimiter=' ',usecols=0)
|
|
29
|
-
b=np.loadtxt('pre10.dat',delimiter=' ',usecols=3)
|
|
30
|
-
c=np.loadtxt('pre10.dat',delimiter=' ',usecols=1)
|
|
31
|
-
d=np.loadtxt('pre10.dat',delimiter=' ',usecols=2)
|
|
32
|
-
a=zscore(a)
|
|
33
|
-
c=zscore(c)
|
|
34
|
-
d=zscore(d)
|
|
35
|
-
signals = np.zeros((N,T))
|
|
36
|
-
for i in range(N):
|
|
37
|
-
signals[i] = a[i]
|
|
38
|
-
sig2=np.zeros((N,T))
|
|
39
|
-
for i in range(N):
|
|
40
|
-
sig2[i] = c[i]
|
|
41
|
-
sig3=np.zeros((N,T))
|
|
42
|
-
for i in range(N):
|
|
43
|
-
sig3[i] = d[i]
|
|
44
|
-
masks = np.zeros((N, T))
|
|
45
|
-
for i in range(N):
|
|
46
|
-
masks[i] = b[i]
|
|
47
|
-
data = np.zeros((N, T, 3))
|
|
48
|
-
data[:, :, 0] = signals[:]
|
|
49
|
-
data[:, :, 1] = sig2[:]
|
|
50
|
-
data[:, :, 2] = sig3[:]
|
|
51
|
-
target = np.zeros((N,T))
|
|
52
|
-
for i in range(N):
|
|
53
|
-
target[i]=b[i]
|
|
54
|
-
return (data, target)
|
|
55
|
-
|
|
56
|
-
def read_test(N=10, T=200):
|
|
57
|
-
a=np.loadtxt('pre14.dat',delimiter=' ',usecols=0)
|
|
58
|
-
c=np.loadtxt('pre14.dat',delimiter=' ',usecols=1)
|
|
59
|
-
d=np.loadtxt('pre14.dat',delimiter=' ',usecols=2)
|
|
60
|
-
a=zscore(a)
|
|
61
|
-
c=zscore(c)
|
|
62
|
-
d=zscore(d)
|
|
63
|
-
signals = np.zeros((N,T))
|
|
64
|
-
for i in range(N):
|
|
65
|
-
signals[i] = a[i]
|
|
66
|
-
sig2=np.zeros((N,T))
|
|
67
|
-
for i in range(N):
|
|
68
|
-
sig2[i] = c[i]
|
|
69
|
-
sig3=np.zeros((N,T))
|
|
70
|
-
for i in range(N):
|
|
71
|
-
sig3[i] = d[i]
|
|
72
|
-
data = np.zeros((N, T, 3))
|
|
73
|
-
data[:, :, 0] = signals[:]
|
|
74
|
-
data[:, :, 1] = sig2[:]
|
|
75
|
-
data[:, :, 2] = sig3[:]
|
|
76
|
-
return (data)
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
'''
|
|
81
|
-
モデルファイル用設定
|
|
82
|
-
'''
|
|
83
|
-
MODEL_DIR = os.path.join(os.path.dirname(__file__), 'model')
|
|
84
|
-
|
|
85
|
-
if os.path.exists(MODEL_DIR) is False:
|
|
86
|
-
os.mkdir(MODEL_DIR)
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
'''
|
|
91
|
-
データの読み取り
|
|
92
|
-
'''
|
|
93
|
-
N = 13999
|
|
94
|
-
T = 200
|
|
95
|
-
maxlen = T
|
|
96
|
-
|
|
97
|
-
X, Y = read(N=N, T=T)
|
|
98
|
-
X_test=read_test(N=N,T=T)
|
|
99
|
-
N_train = int(N * 0.9)
|
|
100
|
-
N_validation = N - N_train
|
|
101
|
-
|
|
102
|
-
'''
|
|
103
|
-
モデル設定
|
|
104
|
-
'''
|
|
105
|
-
n_in = len(X[0][0]) # 2
|
|
106
|
-
n_hidden = 100
|
|
107
|
-
n_out = len(Y[0]) # 1
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
def weight_variable(shape, name=None):
|
|
111
|
-
return np.random.normal(scale=.01, size=shape)
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
early_stopping = EarlyStopping(monitor='loss', patience=100, verbose=1)
|
|
115
|
-
|
|
116
|
-
kernel=weight_variable
|
|
117
|
-
kernel2=weight_variable
|
|
118
|
-
|
|
119
|
-
model = Sequential()
|
|
120
|
-
model.add(GRU(n_hidden,
|
|
121
|
-
#kernel_initializer=kernel,
|
|
122
|
-
input_shape=(maxlen, n_in)))
|
|
123
|
-
model.add(Dense(n_out))
|
|
124
|
-
model.add(Activation('linear'))
|
|
125
|
-
|
|
126
|
-
optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999)
|
|
127
|
-
model.compile(loss='mean_squared_error',
|
|
128
|
-
optimizer=optimizer)
|
|
129
|
-
|
|
130
|
-
'''
|
|
131
|
-
モデル学習
|
|
132
|
-
'''
|
|
133
|
-
|
|
134
|
-
epochs = 1000
|
|
135
|
-
batch_size = 100
|
|
136
|
-
|
|
137
|
-
hist = model.fit(X, Y,
|
|
138
|
-
batch_size=batch_size,
|
|
139
|
-
epochs=epochs,
|
|
140
|
-
callbacks=[early_stopping])
|
|
141
|
-
|
|
142
|
-
model.save(MODEL_DIR+'/model_relu.hdf5')
|
|
143
|
-
print('Model saved')
|
|
144
|
-
|
|
145
|
-
'''
|
|
146
|
-
学習の進み具合を可視化
|
|
147
|
-
'''
|
|
148
|
-
|
|
149
|
-
loss = hist.history['loss']
|
|
150
|
-
|
|
151
|
-
plt.figure(1)
|
|
152
|
-
plt.rc('font', family='serif')
|
|
153
|
-
plt.plot(range(len(loss)), loss, label='loss', color='black')
|
|
154
|
-
plt.xlabel('epochs')
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
predicted = model.predict(X_test)
|
|
159
|
-
A = np.loadtxt('pre14.dat',delimiter=' ',usecols=0)
|
|
160
|
-
A=zscore(A)
|
|
161
|
-
|
|
162
|
-
plt.figure(2)
|
|
163
|
-
plt.ylim([-3, 3])
|
|
164
|
-
plt.plot(A, color='#aaaaaa')
|
|
165
|
-
#plt.plot(B,color='black' )
|
|
166
|
-
plt.plot(predicted, color='red')
|
|
167
|
-
plt.show()
|
|
168
|
-
#plt.savefig(__file__ + '.eps')
|
|
169
|
-
|
|
170
|
-
```
|
|
171
|
-
|
|
172
|
-
###試したこと
|
|
173
|
-
https://teratail.com/questions/105989 で回答を頂いたように，
|
|
174
|
-
```
|
|
175
|
-
dat_paths = glob.glob("pre*.dat")
|
|
176
|
-
a = [np.loadtxt(dat_path,delimiter=' ',usecols=0)for dat_path in dat_paths]
|
|
177
|
-
b = [np.loadtxt(dat_path,delimiter=' ',usecols=3)for dat_path in dat_paths]
|
|
178
|
-
c = [np.loadtxt(dat_path,delimiter=' ',usecols=1)for dat_path in dat_paths]
|
|
179
|
-
d = [np.loadtxt(dat_path,delimiter=' ',usecols=2)for dat_path in dat_paths]
|
|
180
|
-
```
|
|
181
|
-
上記のように読み込み関数を書き換えたのですが，これだけではそれ以降の処理でエラーが出るのと，なにより複数サンプル学習させる方法として正しいのかわかりません．
|
|
182
|
-
ちなみに，他にも少し書き換えているのでコードの全容を載せておきます
|
|
183
|
-
```
|
|
184
|
-
import numpy as np
|
|
185
|
-
import matplotlib.pyplot as plt
|
|
186
|
-
from keras.models import Sequential
|
|
187
|
-
from keras.layers.core import Dense, Activation
|
|
188
|
-
from keras.layers.recurrent import GRU
|
|
189
|
-
from keras.optimizers import Adam
|
|
190
|
-
from keras.callbacks import EarlyStopping
|
|
191
|
-
from sklearn.model_selection import train_test_split
|
|
192
|
-
from sklearn.utils import shuffle
|
|
193
|
-
import os
|
|
194
|
-
import glob
|
|
195
|
-
|
|
196
|
-
|
|
197
|
-
np.random.seed(0)
|
|
198
|
-
|
|
199
|
-
def zscore(x, axis = None):
|
|
200
|
-
xmean = x.mean(axis=axis, keepdims=True)
|
|
201
|
-
xstd = np.std(x, axis=axis, keepdims=True)
|
|
202
|
-
zscore = (x-xmean)/xstd
|
|
203
|
-
return zscore
|
|
204
|
-
|
|
205
|
-
def read(N=10, T=200):
|
|
206
|
-
dat_paths = glob.glob("pre*.dat")
|
|
207
|
-
A = [np.loadtxt(dat_path,delimiter=' ',usecols=0)for dat_path in dat_paths]
|
|
208
|
-
b = [np.loadtxt(dat_path,delimiter=' ',usecols=3)for dat_path in dat_paths]
|
|
209
|
-
C = [np.loadtxt(dat_path,delimiter=' ',usecols=1)for dat_path in dat_paths]
|
|
210
|
-
D = [np.loadtxt(dat_path,delimiter=' ',usecols=2)for dat_path in dat_paths]
|
|
211
|
-
#a=np.loadtxt('*.dat',delimiter=' ',usecols=0)
|
|
212
|
-
#b=np.loadtxt('*.dat',delimiter=' ',usecols=3)
|
|
213
|
-
#c=np.loadtxt('*.dat',delimiter=' ',usecols=1)
|
|
214
|
-
#d=np.loadtxt('*.dat',delimiter=' ',usecols=2)
|
|
215
|
-
|
|
216
|
-
signals = np.zeros((N,T))
|
|
217
|
-
for i in range(N):
|
|
218
|
-
signals[i] = a[i]
|
|
219
|
-
sig2=np.zeros((N,T))
|
|
220
|
-
for i in range(N):
|
|
221
|
-
sig2[i] = c[i]
|
|
222
|
-
sig3=np.zeros((N,T))
|
|
223
|
-
for i in range(N):
|
|
224
|
-
sig3[i] = d[i]
|
|
225
|
-
masks = np.zeros((N, T))
|
|
226
|
-
for i in range(N):
|
|
227
|
-
masks[i] = b[i]
|
|
228
|
-
data = np.zeros((N, T, 3))
|
|
229
|
-
data[:, :, 0] = signals[:]
|
|
230
|
-
data[:, :, 1] = sig2[:]
|
|
231
|
-
data[:, :, 2] = sig3[:]
|
|
232
|
-
target = np.zeros((N,T))
|
|
233
|
-
for i in range(N):
|
|
234
|
-
target[i]=b[i]
|
|
235
|
-
return (data, target)
|
|
236
|
-
|
|
237
|
-
|
|
238
|
-
'''
|
|
239
|
-
モデルファイル用設定
|
|
240
|
-
'''
|
|
241
|
-
MODEL_DIR = os.path.join(os.path.dirname(__file__), 'model')
|
|
242
|
-
|
|
243
|
-
if os.path.exists(MODEL_DIR) is False:
|
|
244
|
-
os.mkdir(MODEL_DIR)
|
|
245
|
-
|
|
246
|
-
|
|
247
|
-
|
|
248
|
-
'''
|
|
249
|
-
データの読み取り
|
|
250
|
-
'''
|
|
251
|
-
N = 500000
|
|
252
|
-
T = 200
|
|
253
|
-
maxlen = T
|
|
254
|
-
|
|
255
|
-
X, Y = read(N=N, T=T)
|
|
256
|
-
N_train = int(N * 0.9)
|
|
257
|
-
N_validation = N - N_train
|
|
258
|
-
|
|
259
|
-
X_train,X_test,Y_train,Y_test=\
|
|
260
|
-
train_test_split(X,Y,train_size=train_size)
|
|
261
|
-
X_train,X_validation,Y_train,Y_validation= \
|
|
262
|
-
train_test_split(X_train,Y_train,train_size=N_validation)
|
|
263
|
-
|
|
264
|
-
'''
|
|
265
|
-
モデル設定
|
|
266
|
-
'''
|
|
267
|
-
n_in = len(X[0][0]) # 2
|
|
268
|
-
n_hidden = 100
|
|
269
|
-
n_out = len(Y[0]) # 1
|
|
270
|
-
|
|
271
|
-
|
|
272
|
-
def weight_variable(shape, name=None):
|
|
273
|
-
return np.random.normal(scale=.01, size=shape)
|
|
274
|
-
|
|
275
|
-
|
|
276
|
-
early_stopping = EarlyStopping(monitor='loss', patience=100, verbose=1)
|
|
277
|
-
|
|
278
|
-
kernel=weight_variable
|
|
279
|
-
kernel2=weight_variable
|
|
280
|
-
|
|
281
|
-
model = Sequential()
|
|
282
|
-
model.add(GRU(n_hidden,
|
|
283
|
-
#kernel_initializer=kernel,
|
|
284
|
-
input_shape=(maxlen, n_in)))
|
|
285
|
-
model.add(Dense(n_out))
|
|
286
|
-
model.add(Activation('linear'))
|
|
287
|
-
|
|
288
|
-
optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999)
|
|
289
|
-
model.compile(loss='mean_squared_error',
|
|
290
|
-
optimizer=optimizer)
|
|
291
|
-
|
|
292
|
-
'''
|
|
293
|
-
モデル学習
|
|
294
|
-
'''
|
|
295
|
-
|
|
296
|
-
epochs = 1000
|
|
297
|
-
batch_size = 100
|
|
298
|
-
|
|
299
|
-
hist = model.fit(X_train, Y_train,
|
|
300
|
-
batch_size=batch_size,
|
|
301
|
-
epochs=epochs,
|
|
302
|
-
validation_data=(X_validation,Y_validation),
|
|
303
|
-
callbacks=[early_stopping])
|
|
304
|
-
|
|
305
|
-
model.save(MODEL_DIR+'/model_GRU.hdf5')
|
|
306
|
-
print('Model saved')
|
|
307
|
-
|
|
308
|
-
'''
|
|
309
|
-
学習の進み具合を可視化
|
|
310
|
-
'''
|
|
311
|
-
|
|
312
|
-
|
|
313
|
-
val_loss=hist.history['val_loss']
|
|
314
|
-
val_acc=hist.history['val_acc']
|
|
315
|
-
|
|
316
|
-
loss_and_metrics=model.evaluate(X_test,Y_test)
|
|
317
|
-
print(loss_and_metrics)
|
|
318
|
-
|
|
319
|
-
plt.figure(1)
|
|
320
|
-
plt.plot(range(len(val_acc)),val_acc,label='acc',color='red')
|
|
321
|
-
plt.xlabel('epochs')
|
|
322
|
-
|
|
323
|
-
plt.figure(2)
|
|
324
|
-
plt.plot(range(len(val_loss)),val_loss,label='loss',color='red')
|
|
325
|
-
plt.xlabel('epochs')
|
|
326
|
-
|
|
327
|
-
plt.show()
|
|
328
|
-
|
|
329
|
-
```
|
1
タグの編集
title
CHANGED
|
File without changes
|
body
CHANGED
|
File without changes
|