-
Notifications
You must be signed in to change notification settings - Fork 0
/
gcn_lstm.py
157 lines (131 loc) · 6.08 KB
/
gcn_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
import argparse
import numpy as np
import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt
import scipy.sparse as sp
import time
import pickle
import os
import math
from more_itertools import sample
from yaml import parse
from sklearn.metrics import mean_squared_error
from sklearn.preprocessing import MinMaxScaler
from scipy.sparse.linalg.eigen.arpack import eigsh, ArpackNoConvergence
from scipy import stats
from tensorflow.keras.layers import Input, Dropout, Dense, Layer, Reshape, Lambda, LSTM, Activation, Add, BatchNormalization
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam
from model import GraphConvolution
from utils import *
if __name__ == "__main__" :
parser = argparse.ArgumentParser()
parser.add_argument('-v', '--version', default="v1_1")
parser.add_argument('--folder', default="V1")
parser.add_argument('-f','--nfeat', type=int, default=90)
parser.add_argument('-s','--seed', type=int, default=0)
parser.add_argument('--support', default=1)
parser.add_argument('--house_size', type = int, default=1600)
args = parser.parse_args()
tf.random.set_seed(args.seed)
wait = 0
preds = None
best_val_mse = 9999999
NB_EPOCH = 1000
PATIENCE = 50
X = np.load(args.folder + "/feature_matrix_"+ args.version+".npy")
X = MinMaxScaler().fit_transform(X)
D = sp.load_npz(args.folder + "/adjacency_matrix_" + args.version + ".npz")
S = sp.load_npz(args.folder + "/status_matrix_" + args.version + ".npz")
y = np.load(args.folder + "/label_nodes_"+ args.version + ".npy")/100000
A = D.multiply(S)
A_ = preprocess_adj(A, True)
graph = [X, A_]
G = [Input(shape=(None,None), sparse=True)]
y_train, y_val, y_test, idx_train, idx_val, idx_test, train_mask = get_splits(y)
whole_house = X.shape[0] #number of houses available
house_size = args.house_size #number of house sampled for each window time
seq_len = int(whole_house/house_size) #number of window time
#Create Model
X_in = Input(shape=(X.shape[1],))
H = X_in
H = GraphConvolution(args.nfeat, args.support, activation='relu')([H]+G)
H = GraphConvolution(args.nfeat, args.support, activation='relu')([H]+G)
H = Dropout(0.5)(H)
H = Lambda(lambda H: tf.reshape(H, shape=(-1, H.shape[-1]), name="reshape_layer"))(H)
seq_list = []
for i in range(seq_len):
seq_list.append(tf.gather(params=H, indices=range(i*house_size, (i+1)*house_size)))
sequence = tf.stack(seq_list, 1)
L,S,R= LSTM(units=90,stateful=True, return_sequences=True,return_state=True)(sequence)
L = Lambda(lambda x : tf.keras.backend.reshape(x, (-1, args.nfeat)))(L)
Y = Dense(1, activation = 'linear')(L)
model = Model(inputs = [X_in]+G, outputs = Y)
model.compile(optimizer=Adam(lr=0.001), loss = tf.keras.losses.MeanSquaredError(), metrics=[tf.keras.metrics.RootMeanSquaredError(),'mae'])
#Loss list
all_val_mse = list()
all_train_mse = list()
all_val_mae = list()
all_train_mae = list()
all_val_error = list()
all_train_error = list()
for epoch in range(1, NB_EPOCH+1):
t = time.time()
#Single training iteration (we mask nodes without labels for loss calculation)
history = model.fit(graph, y_train, sample_weight=train_mask,
batch_size=A.shape[0], epochs=1, shuffle=False, verbose=0)
#Predict on full dataset
preds = model.predict(graph, batch_size=A.shape[0])
#Train/Test Scores
(Train_mse, Val_mse),(Train_error, Val_error), (Train_mae, Val_mae) = evaluate_preds(
preds, [y_train, y_val], [idx_train, idx_val])
all_train_mse.append(Train_mse)
all_val_mse.append(Val_mse)
all_train_mae.append(Train_mae)
all_val_mae.append(Val_mae)
all_train_error.append(Train_error)
all_val_error.append(Val_error)
print("Epoch: {:04d}".format(epoch),
"train_RMSE= {:.4f}".format(Train_mse),
"val_RMSE= {:.4f}".format(Val_mse),
"train_MAE= {:.4f}".format(Train_mae),
"val_MAE= {:.4f}".format(Val_mae),
"train_GER= {:.4f}".format(Train_error),
"val_GER= {:.4f}".format(Val_error),
"time= {:.4f}".format(time.time() - t))
#Early Stopping
if Val_mse < best_val_mse:
best_val_mse = Val_mse
wait = 0
else:
print("Current Best Test MSE:", best_val_mse)
if wait >= PATIENCE:
print("Epoch {}: Early stopping".format(epoch))
break
wait += 1
(Test_mse), (Test_error), (Test_mae) = evaluate_preds(preds, [y_test], [idx_test])
#pearson correlation between predicted value vs ground truth
pred = preds[idx_test]
pred = np.reshape(pred, len(pred))
test_label = y_test[idx_test]
test_label = np.reshape(test_label, len(test_label))
pearson_correlation = stats.pearsonr(pred, test_label)
print('Test Data MSE:', Test_mse)
print('Test Data MAE:', Test_mae)
print('Pearson Correlation: ', pearson_correlation)
loss_data = [all_train_mse, all_val_mse, all_train_mae, all_val_mae, all_train_error, all_val_error]
loss_dict = ['Train_Loss(RMSE)', 'Val_Loss(RMSE)', 'Train_MAE', 'Val_MAE', 'Train_GER', 'Val_GER']
zipObj = zip(loss_dict, loss_data)
loss_history = dict(zipObj)
with open(args.folder+'/saved_model/gcn_lstm_loss_{}'.format(args.version),'wb') as gcn_loss:
pickle.dump(loss_history, gcn_loss, protocol=pickle.HIGHEST_PROTOCOL)
model.save(args.folder+'/saved_model/gcn_lstm_model_{}'.format(args.version))
plt.plot(range(0,len(all_val_mse)), all_val_mse, label = "Val Loss")
plt.plot(range(0,len(all_train_mse)), all_train_mse, label = "Train Loss")
plt.title('Model Loss')
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.ylim([0,6])
plt.legend()
plt.savefig(args.folder+'/saved_model/plot_gcn_lstm_loss_{}'.format(args.version))