-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathbs.py
200 lines (166 loc) · 7.84 KB
/
bs.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
#input X,C, f_x,f_c,adjlist_s,adjlist_c
#output X_t
# (x,c)->z_x /// ()
# input_dim, gcn_output_dim, output_dim, activation, rnn_units,featureless
import tensorflow as tf
from BikeS.dataprocess import loadinput, MinMaxNormalization,station_file_num
import BikeS.Egcn as Egcn
from tensorflow.keras import optimizers
from BikeS.Metrics import pred_MSE, news_MSE
import csv
import pandas as pd
import datetime
import numpy as np
city = 'wash'
def mask(metric):
m_one = tf.ones_like(metric)
m_zero = tf.zeros_like(metric)
metric_onehot = tf.where(metric > 0, x=m_one, y=m_zero)
return metric_onehot
# *********** hyp-para*************
path,cluster_adj='',''
if city=='nyc':
path = 'D:/my/dataset/citibike/data/day/'
cluster_adj = pd.read_csv('D:/my/dataset/citibike/data/cluster.csv',header=None).values
cluster_feat = pd.read_csv('D:/my/dataset/citibike/data/station_cluster.csv',header=None).values
new_station = pd.read_csv('D:/my/dataset/citibike/data/new_stations.csv',header=None).values
learning_rate = 5e-4
seq_len = 6
c_gcn_output_dim = 64
s_gcn_output_dim = 128
activation = tf.nn.tanh
rnn_units = 128
loop_num = 5
lamda_reg = 0.001
else:
path = 'D:/my/dataset/capitalbike/data/day/'
cluster_adj =pd.read_csv('D:/my/dataset\capitalbike/data/cluster.csv',header=None).values
cluster_feat = pd.read_csv('D:/my/dataset/capitalbike/data/station_cluster.csv', header=None).values
new_station = pd.read_csv('D:/my/dataset/capitalbike/data/new_stations.csv',header=None).values
learning_rate = 5e-4
seq_len = 6
c_gcn_output_dim = 32
s_gcn_output_dim = 64
activation = tf.nn.tanh
rnn_units = 128
loop_num = 5
lamda_reg = 0.001
station_num, timestamp =station_file_num(path)
cluster_num = cluster_adj.shape[1]
# print(station_num)
output_dim = station_num
# path for result
w_filename = 'D:\\my\\mypaper\\BIKEsharing\\results\\'+city+'model_res_end_egvae.csv'
ns_w_filename = 'D:\\my\\mypaper\\BIKEsharing\\results\\'+city+'WEIGHT_res_end_egvae.csv'
# ----Model initial-------
c_hid_layer_list = [cluster_num, c_gcn_output_dim]
s_hid_layer_list = [station_num, s_gcn_output_dim]
# ------inputs load ----------
def selectModel(m):
if m == 'GCNGRU':
MODEL = Egcn.GCNGRU(station_num, s_gcn_output_dim, output_dim, activation, rnn_units, seq_len)
elif m=='GRU':
MODEL = Egcn.GRU(output_dim,rnn_units,seq_len)
elif m == 'CGCNGRU':
MODEL = Egcn.C_GCNGRU(c_hid_layer_list, s_hid_layer_list, output_dim, activation, rnn_units, seq_len)
elif m == 'GCNVAE':
MODEL = Egcn.GCNVAE(c_hid_layer_list, s_hid_layer_list, output_dim, rnn_units, seq_len)
elif m == 'EGCN_VAE':
MODEL = Egcn.EGCNVAE(c_hid_layer_list, s_hid_layer_list, rnn_units, seq_len, output_dim)
elif m=="MG_VAE":
MODEL = Egcn.MG_VAE(c_hid_layer_list,s_hid_layer_list,cluster_num,rnn_units,seq_len,output_dim)
elif m == 'EGCN':
MODEL = Egcn.EGCN(c_hid_layer_list, s_hid_layer_list, output_dim, rnn_units, seq_len)
return MODEL
whole_dataset_len = timestamp-seq_len
print('whole_dataset_len :{}-----------learning rate:{}'.format(whole_dataset_len,learning_rate))
cluster_adj = tf.expand_dims(tf.convert_to_tensor(cluster_adj,dtype=tf.float32),axis=0)
cluster_feat = tf.convert_to_tensor(cluster_feat,dtype=tf.float32)
# Model initialization
# models=['GRU','GCNGRU','CGCNGRU','EGCN','GCNVAE','EGCN_VAE']
models=['EGCN_VAE']
optimizer = optimizers.Adam(lr=learning_rate)
for m in models:
MODEL = selectModel(m) # initialization
NS_RES_writer = csv.writer(open(ns_w_filename, 'w', newline=''))
# result writers
GCNCRU_RES_writer = csv.writer(open(w_filename, 'a', newline=''))
cur_time = datetime.datetime.now()
GCNCRU_RES_writer.writerow([learning_rate,lamda_reg, seq_len, s_gcn_output_dim, rnn_units, m,'0.01w',cur_time])
GCNCRU_RES_writer.writerow(['date', 'MAE', 'RMSE', 'N_MAE', 'N_RMSE','O_MAE','O_RMSE'])
# NS_RES_writer.writerow([m, cur_time])
print('MODEL_NAME:',m)
# record losses for each epoch
EPOCH = []
start_NORM = MinMaxNormalization()
end_NORM = MinMaxNormalization()
inputs = loadinput(city, seq_len, start_NORM, end_NORM, loop_num)
for i in range(loop_num):
EPOCH.append(0)
loop = 0
epoch = 0
loss=0
while epoch < whole_dataset_len - 1:
inputs_list = inputs.__next__()
adj_list = tf.expand_dims(tf.convert_to_tensor(inputs_list[0], dtype=tf.float32), axis=0)
x_list = tf.expand_dims(tf.convert_to_tensor(inputs_list[1], dtype=tf.float32), axis=0)
label = tf.convert_to_tensor(inputs_list[-1], dtype=tf.float32)
time_feat = tf.expand_dims(tf.convert_to_tensor(inputs_list[2], dtype=tf.float32), axis=0)
label_mask = tf.expand_dims(mask(x_list[0, 0, :, 0]), axis=-1)
# label_mask = tf.squeeze(mask(label),axis=0)
cluster_feat=label_mask*cluster_feat #[s_num,1]*[s_num,c_num]
if m == 'GCNGRU':
INPUTS = (adj_list, x_list, time_feat, label)
elif m=='GRU':
INPUTS = (x_list,time_feat,label)
else:
INPUTS = (adj_list, x_list, cluster_adj, cluster_feat, time_feat, label)
if epoch + 1 >= whole_dataset_len * 0.85 and loop < loop_num:
if loop < loop_num - 1:
epoch = 0
loop += 1
# train
if epoch < whole_dataset_len * 0.85 and loop < loop_num:
# GCNGRU_
print('--Train-epoch:{}-sample:{}---:'.format(loop, epoch), '----date:', inputs_list[-2])
with tf.GradientTape() as tape:
l2_reg = tf.reduce_mean([tf.nn.l2_loss(v) for v in MODEL.trainable_variables])
if 'VAE' in m:
outputs,_= MODEL(INPUTS)
loss= Egcn.vae_lossfunction(label, outputs[0], outputs[1:], label_mask, loop) + lamda_reg * l2_reg
# loss+=loss_i
MAE, RMSE = pred_MSE(label.numpy(), outputs[0].numpy(), start_NORM, label_mask.numpy())
else: # GCNGRU C_GCNGRU EGCN
outputs = MODEL(INPUTS)
loss= Egcn.lossfunction(label, outputs, label_mask) + lamda_reg * l2_reg
# loss+=loss_i
MAE, RMSE = pred_MSE(label.numpy(), outputs.numpy(), start_NORM, label_mask.numpy())
print('---------loss:', float(loss), '-------RMSE:', float(RMSE))
EPOCH[loop] += loss
# if (epoch + 1) % 5 == 0:
# loss = loss / 5
grads = tape.gradient(loss, MODEL.trainable_variables)
optimizer.apply_gradients(zip(grads, MODEL.trainable_variables))
# print('----Train output----')
# loss = 0
# validation
elif epoch < whole_dataset_len*0.9:
output = MODEL.predict(INPUTS)
# output = GCNGRU.predict(X_inputs)
MAE,RMSE = pred_MSE(label.numpy(),output[0],start_NORM,label_mask.numpy())
print('--------Validation--sample:{}---date:{}---MAE:{}--RMSE:{}'.format(epoch,inputs_list[-2],MAE,RMSE))
# test
else:
# new station for each time
output= MODEL.predict(INPUTS)
# output = GCNGRU.predict(X_inputs)
MAE, RMSE = pred_MSE(label.numpy(), output[0], start_NORM, label_mask.numpy())
rate,n_mae,n_rmse,o_mae,o_rmse = news_MSE(label.numpy(), output[0], start_NORM, new_station, label_mask.numpy())
print('----------TEST----date:{}---MAE:{}----RMSE:{}'.format(inputs_list[-2], MAE, RMSE))
GCNCRU_RES_writer.writerow([inputs_list[-2], MAE, RMSE, n_mae, n_rmse,o_mae,o_rmse])
sc_weight =np.expand_dims(x_list[0, 0, :, 0],axis=-1)* output[-1][0]
NS_RES_writer.writerow(inputs_list[-2])
NS_RES_writer.writerows(sc_weight)
epoch += 1
print(EPOCH)
# test