-
Notifications
You must be signed in to change notification settings - Fork 30
/
ResLSTM.py
179 lines (155 loc) · 7.55 KB
/
ResLSTM.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
from numpy.random import seed
seed(1)
from tensorflow import set_random_seed
set_random_seed(2)
import numpy as np
np.set_printoptions(threshold=np.inf)
import time, os
import keras
keras.backend.set_image_data_format('channels_last')
from keras.layers import *
from keras.models import *
from keras.utils.vis_utils import plot_model #visualize model
from keras.models import load_model
from keras.optimizers import Adam
from metrics import evaluate_performance
from load_data import Get_All_Data
# os.chdir('D:/论文2/upload to GitHub/')
os.environ["PATH"] += os.pathsep + 'E:/Program Files (x86)/Graphviz2.38/bin' #used for visualizing the model
global_start_time = time.time()
def Unit(x, filters, pool=False):
res = x
if pool:
x = MaxPooling2D(pool_size=(2, 2), padding="same")(x)
res = Conv2D(filters=filters, kernel_size=[1, 1], strides=(2, 2), padding="same")(res)
out = BatchNormalization()(x)
out = Activation("relu")(out)
out = Conv2D(filters=filters, kernel_size=[3, 3], strides=[1, 1], padding="same")(out)
out = BatchNormalization()(out)
out = Activation("relu")(out)
out = Conv2D(filters=filters, kernel_size=[3, 3], strides=[1, 1], padding="same")(out)
out = keras.layers.add([res, out])
return out
def attention_3d_block(inputs,timesteps):
#input_dim = int(inputs.shape[2])
a = Permute((2, 1))(inputs)
a = Dense(timesteps, activation='linear')(a)
a_probs = Permute((2, 1))(a)
#output_attention_mul = merge([inputs, a_probs], name='attention_mul', mode='mul')
output_attention_mul = multiply([inputs, a_probs])
return output_attention_mul
# Define the model
def multi_input_model(time_lag):
"""build multi input model构建多输入模型"""
input1_ = Input(shape=(276, time_lag-1, 3), name='input1')
input2_ = Input(shape=(276, time_lag-1, 3), name='input2')
input3_ = Input(shape=(276, time_lag-1, 1), name='input3')
input4_ = Input(shape=(11, time_lag-1, 1), name='input4')
#first input
x1 = Conv2D(filters=32, kernel_size=[3, 3], strides=[1, 1], padding="same")(input1_)
x1 = Unit(x1, 32)
x1 = Unit(x1, 64, pool=True)
x1 = Flatten()(x1)
x1 = Dense(276)(x1)
# second input
x2 = Conv2D(filters=32, kernel_size=[3, 3], strides=[1, 1], padding="same")(input2_)
x2 = Unit(x2, 32)
x2 = Unit(x2, 64, pool=True)
x2 = Flatten()(x2)
x2 = Dense(276)(x2)
# third input
x3 = Conv2D(filters=32, kernel_size=[3, 3], strides=[1, 1], padding="same")(input3_)
x3 = Unit(x3, 32)
x3 = Unit(x3, 64, pool=True)
x3 = Flatten()(x3)
x3 = Dense(276)(x3)
# fourth input
x4 = Flatten()(input4_)
x4 = Dense(276)(x4)
x4 = Reshape(target_shape=(276, 1))(x4)
x4 = LSTM(128, return_sequences=True, input_shape=(276, 1))(x4)
x4 = LSTM(276, return_sequences=False)(x4)
x4 = Dense(276)(x4)
out = keras.layers.add([x1, x2, x3, x4])
out = Reshape(target_shape=(276, 1))(out)
out = LSTM(128, return_sequences=True,input_shape=(276, 1))(out)
out = attention_3d_block(out, 276)#shape of the output is（276，128）
out = Flatten()(out)
out = Dense(276)(out)
model = Model(inputs=[input1_, input2_, input3_,input4_], outputs=[out]) #[input1_, input2_, input3_]
return model
def build_model(X_train_1,X_train_2,X_train_3,X_train_4,Y_train,X_test_1,X_test_2,X_test_3,X_test_4,Y_test,\
Y_test_original,batch_size,epochs,a,time_lag):
X_train_1 = X_train_1.reshape(X_train_1.shape[0], 276, time_lag-1, 3)
X_train_2 = X_train_2.reshape(X_train_2.shape[0], 276, time_lag-1, 3)
X_train_3 = X_train_3.reshape(X_train_3.shape[0], 276, time_lag-1, 1)
X_train_4 = X_train_4.reshape(X_train_4.shape[0], 11, time_lag-1, 1)
Y_train = Y_train.reshape(Y_train.shape[0], 276)
X_test_1 = X_test_1.reshape(X_test_1.shape[0], 276, time_lag-1, 3)
X_test_2 = X_test_2.reshape(X_test_2.shape[0], 276, time_lag-1, 3)
X_test_3 = X_test_3.reshape(X_test_3.shape[0], 276, time_lag-1, 1)
X_test_4 = X_test_4.reshape(X_test_4.shape[0], 11, time_lag-1, 1)
Y_test = Y_test.reshape(Y_test.shape[0], 276)
if epochs == 50:
model = multi_input_model(time_lag)
model.compile(optimizer=Adam(), loss='mse', metrics=['mse'])
model.fit([X_train_1, X_train_2, X_train_3, X_train_4], Y_train, batch_size=batch_size, epochs=epochs, verbose=2, shuffle=False)#, validation_split=0.05
output = model.predict([X_test_1, X_test_2, X_test_3, X_test_4], batch_size=batch_size)
else:
# train models every 10 epoches
model = load_model('testresult/'+str(epochs-10)+'-model-with-graph.h5')
model.fit([X_train_1, X_train_2, X_train_3, X_train_4], Y_train, batch_size=batch_size, epochs=10, verbose=2, shuffle=False)# , validation_split=0.05
output = model.predict([X_test_1, X_test_2, X_test_3, X_test_4], batch_size=batch_size)
#rescale the output of this model将输出进行反归一化
predictions = np.zeros((output.shape[0], output.shape[1]))
for i in range(len(predictions)):
for j in range(len(predictions[0])):
predictions[i, j] = round(output[i, j]*a, 0)
if predictions[i, j] < 0:
predictions[i, j] = 0
RMSE,R2,MAE,WMAPE=evaluate_performance(Y_test_original,predictions)
#visualize the model structure
plot_model(model, to_file='model.png', show_shapes=True)
#print(model.summary())
return model,Y_test_original,predictions,RMSE,R2,MAE,WMAPE
def Save_Data(path,model,Y_test_original,predictions,RMSE,R2,MAE,WMAPE,Run_epoch):
print(Run_epoch)
RMSE_ALL=[]
R2_ALL=[]
MAE_ALL=[]
WMAPE_ALL=[]
Average_train_time=[]
RMSE_ALL.append(RMSE)
R2_ALL.append(R2)
MAE_ALL.append(MAE)
WMAPE_ALL.append(WMAPE)
model.save(path+str(Run_epoch)+'-model-with-graph.h5')
np.savetxt(path+str(Run_epoch)+'-RMSE_ALL.txt', RMSE_ALL)
np.savetxt(path+str(Run_epoch)+'-R2_ALL.txt', R2_ALL)
np.savetxt(path+str(Run_epoch)+'-MAE_ALL.txt', MAE_ALL)
np.savetxt(path+str(Run_epoch)+'-WMAPE_ALL.txt', WMAPE_ALL)
with open(path+str(Run_epoch)+'-predictions.csv', 'w') as file:
predictions = predictions.tolist()
for i in range(len(predictions)):
file.write(str(predictions[i]).replace("'", "").replace("[", "").replace("]", "")+"\n")
with open(path+str(Run_epoch)+'-Y_test_original.csv', 'w') as file:
Y_test_original = Y_test_original.tolist()
for i in range(len(Y_test_original)):
file.write(str(Y_test_original[i]).replace("'", "").replace("[", "").replace("]", "")+"\n")
duration_time = time.time() - global_start_time
Average_train_time.append(duration_time)
np.savetxt(path+str(Run_epoch)+'-Average_train_time.txt', Average_train_time)
print('total training time(s):', duration_time)
X_train_1,Y_train,X_test_1,Y_test,Y_test_original,a,b,X_train_2,X_test_2,X_train_3,X_test_3,X_train_4,X_test_4=\
Get_All_Data(TG=15, time_lag=6, TG_in_one_day=72, forecast_day_number=5, TG_in_one_week=360)
Run_epoch = 50 # first training 50 epoch, and then add 10 epoch every time 初始训练epoch，以后每次加10，运行15次
for i in range(15):
model, Y_test_original, predictions, RMSE, R2, MAE, WMAPE = build_model(X_train_1,X_train_2,X_train_3,X_train_4,Y_train,X_test_1,X_test_2,X_test_3,X_test_4,Y_test,\
Y_test_original,batch_size=64,epochs=Run_epoch,a=a,time_lag=6)
Save_Data("testresult/", model, Y_test_original, predictions, RMSE, R2, MAE, WMAPE, Run_epoch)
Run_epoch += 10
#For Get_All_Data, change parameters referring to this: TG=15, time_lag=6, TG_in_one_day=72, forecast_day_number=5, TG_in_one_week=360
#10min:10,6,108,5,540,eopch=200
#15min:15,6,72,5,360 eopch=140
#30min:30,6,36,5,180 eopch=200
#60min:60,6,18,5,90 eopch=235