-
Notifications
You must be signed in to change notification settings - Fork 1
/
LSTM-ensemble.py
180 lines (146 loc) · 5.69 KB
/
LSTM-ensemble.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
#!/usr/bin/env python
# coding: utf-8
# In[1]:
import keras
import tensorflow as tf
import numpy as np
from matplotlib import pyplot as plt
from scipy.stats import pearsonr
from sklearn.preprocessing import MinMaxScaler, StandardScaler
from keras.models import Sequential, Model
from scipy.stats import pearsonr
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.backend.tensorflow_backend import set_session
from keras import backend as K
from keras.backend import slice
from sklearn.metrics import mean_absolute_error, mean_squared_error
from scipy.stats import pearsonr
from sklearn.linear_model import LinearRegression
from sklearn.model_selection import train_test_split
import time
import pickle
import os
# In[2]:
os.environ["CUDA_VISIBLE_DEVICES"]="1"
# ## Functions
# In[3]:
def load_data(station): ## Station in String
flow = np.load('../usgsflow_'+station+'.npy')
precip = np.load('../NLDAS_precip_'+station+'.npy')
srad = np.load('../NLDAS_srad_'+station+'.npy')
tmax = np.load('../NLDAS_tmax_'+station+'.npy')
y = np.array(flow).reshape(-1, 1)
indx = np.where(y>=0)[0]
# print(precip.shape)
date = np.load('../usgsdate_'+station+'.npy', allow_pickle=True)
x = np.concatenate((precip, srad, tmax), axis=1)
x = x[indx]
y = y[indx]
date = date[indx]
return x, y
def nse(y_pred, y_true):
nse = 1-np.sum((y_pred-y_true)**2)/np.sum((y_true-np.mean(y_true))**2)
return nse
def dataset_ld(x,y,W,L):
obs = x.shape[0]
features = x.shape[1]
a = np.zeros([obs-W-L+1, W, features])
b = np.zeros([obs-W-L+1, 1])
for i in range(obs-W-L+1):
a[i,:,:] = x[i:i+W,:]
b[i,:] = y[i+W+L-1,0]
return a, b
def train_test_pre(x, y):
xtrain = x[:10000]; xtest = x[10000:]
ytrain = y[:10000]; ytest = y[10000:]
xscale = StandardScaler().fit(xtrain)
yscale = StandardScaler().fit(ytrain)
Xtrain = xscale.transform(xtrain); Xtest = xscale.transform(xtest)
Ytrain = yscale.transform(ytrain); Ytest = yscale.transform(ytest)
return Xtrain, Xtest, Ytrain, Ytest, xscale, yscale
def custom_loss(y_true, y_pred):
s1 = K.sum((y_pred-y_true)**2)/K.sum((y_true-K.mean(y_true))**2)
return s1
# In[4]:
def build_model(W,L):
model = Sequential()
model.add(keras.layers.CuDNNLSTM(50, return_sequences=True))
model.add(keras.layers.CuDNNLSTM(50, return_sequences=True))
model.add(keras.layers.CuDNNLSTM(50, return_sequences=False))
model.add(keras.layers.Dropout(0.5))
model.add(keras.layers.Dense(100))
model.add(keras.layers.Dense(1))
return model
# ## Constants
# In[5]:
lr = 0.0005; W=365; L=0;
f = open('../../StationArea.pkl','rb')
areas = pickle.load(f); f.close()
# In[ ]:
# stations = np.load('../station-list.npy')
stations = np.load('../first-stations.npy')
d_nse = np.zeros((15,20)); d_mse = np.zeros((15,20)); d_mae = np.zeros((15,20)); d_r = np.zeros((15,20));
high_mse = np.zeros((15,20))
low_mse = np.zeros((15,20))
i=0
for station in stations:
x, y = load_data(str(station))
area = areas[str(station)]
## Transform to Runoff
y = y*86400*1000/(area*1000*1000)
a_nse = []; a_mse = []; a_r = []; a_mae = []; total_time=0
a_high_mse=[]; a_low_mse=[]
best_nse = -100; model_name='LSTM/'+str(station)+'_LSTM.h5' ## Save the best nse and best model.
Xtrain, Xtest, Ytrain, Ytest, xscale, yscale = train_test_pre(x, y)
X_train, Y_train = dataset_ld(Xtrain, Ytrain, W, L)
X_test, Y_test = dataset_ld(Xtest, Ytest, W, L)
for training_id in range(15):
ensemble_name = 'LSTM/'+str(station)+'_LSTM_'+str(training_id)+'.h5'
model = build_model(W,L)
adam = keras.optimizers.Adam(lr=lr)
# model.compile(loss='mse', optimizer=adam)
model.compile(loss=custom_loss, optimizer=adam)
# X_train, X_val, Y_train, Y_val = train_test_split(X_train, Y_train, test_size=0.25, random_state=7)
# training
start = time.time()
history = model.fit(X_train, Y_train, epochs=150, batch_size=512,
verbose=0, shuffle=True)
run_time = time.time()-start
total_time+=run_time
# testing:
Y_pred = model.predict(X_test)
y_pred = yscale.inverse_transform((Y_pred).reshape(-1, 1))
y_true = yscale.inverse_transform((Y_test).reshape(-1, 1))
NSE = nse(y_pred, y_true);
R = pearsonr(y_pred.flatten(), y_true.flatten())[0]
MSE = mean_squared_error(y_pred, y_true); MAE = mean_absolute_error(y_pred, y_true)
a_nse+=[NSE]; a_mse+=[MSE]; a_r+=[R]; a_mae+=[MAE]
## Save the best nse and best model.
if (NSE>best_nse):
best_nse=NSE; best_model=model; print('better')
model.save_weights(ensemble_name)
## High flow and low flow
ind = np.argwhere(y_true<=np.percentile(y_true, 5))
low_pred = y_pred[ind]
low_true = y_true[ind]
ind = np.argwhere(y_true>=np.percentile(y_true, 95))
high_pred = y_pred[ind]
high_true = y_true[ind]
e = high_pred-high_true; mse = np.mean(np.square(e)); a_high_mse+=[mse]
e = low_pred-low_true; mse = np.mean(np.square(e)); a_low_mse+=[mse]
del model
del adam
print(station,': run time is ', total_time/15, 's')
# print('NSE: ', a_nse, ' R: ', a_r)
# print('MSE: ', a_mse, ' MAE: ', a_mae)
d_nse[:,i]=a_nse; d_mse[:,i]=a_mse; d_mae[:,i]=a_mae; d_r[:,i]=a_r
high_mse[:,i]=a_high_mse; low_mse[:,i]=a_low_mse;
i+=1
# model.save(model_name)
best_model.save_weights(model_name)
# In[ ]:
np.save('LSTM_NSE', d_nse);
np.save('LSTM_MSE', d_mse);
np.save('LSTM_high_mse', high_mse)
np.save('LSTM_low_mse', low_mse)
print('done')