-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_ww.py
157 lines (136 loc) · 6.35 KB
/
model_ww.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 torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torch.optim as optim
import matplotlib.pyplot as plt
class STResNet(nn.Module):
def __init__(self, c_conf=(2, 2, 32, 32), p_conf=(2, 2, 32, 32), t_conf=(2, 2, 32, 32), external_dim=-1, nb_residual_unit=3):
super(STResNet, self).__init__()
'''
C - Temporal Closeness
P - Period
T - Trend
conf = (len_seq, nb_flow, map_height, map_width)
external_dim
'''
self.external_dim = external_dim
self.nb_residual_unit = nb_residual_unit
len_seq, nb_flow, map_height, map_width = c_conf
self.c_conv1 = nn.Conv2d(in_channels=nb_flow * len_seq, out_channels=64, kernel_size=3, padding=1)
self.c_conv2 = nn.Conv2d(in_channels=64, out_channels=nb_flow, kernel_size=3, padding=1)
len_seq, nb_flow, map_height, map_width = p_conf
self.p_conv1 = nn.Conv2d(in_channels=nb_flow * len_seq, out_channels=64, kernel_size=3, padding=1)
self.p_conv2 = nn.Conv2d(in_channels=64, out_channels=nb_flow, kernel_size=3, padding=1)
# len_seq, nb_flow, map_height, map_width = t_conf
# self.t_conv1 = nn.Conv2d(in_channels=nb_flow * len_seq, out_channels=64, kernel_size=3, padding=1)
# self.t_conv2 = nn.Conv2d(in_channels=64, out_channels=nb_flow, kernel_size=3, padding=1)
# self.linear = nn.Linear(in_features=self.external_dim, out_features= 49)
# self.deconv1 = nn.ConvTranspose2d(in_channels=1, out_channels=4, kernel_size=5)
# self.deconv2 = nn.ConvTranspose2d(in_channels=4, out_channels=8, kernel_size=5)
# self.deconv3 = nn.ConvTranspose2d(in_channels=8, out_channels=8, kernel_size=3, stride=2)
# self.deconv4 = nn.ConvTranspose2d(in_channels=8, out_channels=8, kernel_size=5, stride=2)
# self.deconv5 = nn.ConvTranspose2d(in_channels=8, out_channels=4, kernel_size=5, stride=3)
# self.deconv6 = nn.ConvTranspose2d(in_channels=4, out_channels=1, kernel_size=4)
# self.relu = nn.ReLU()
self.nb_flow = nb_flow
self.map_height = map_height
self.map_width = map_width
def _shortcut(self, input, residual):
return torch.add(input, residual)
def _bn_relu_conv(self, nb_filter, ns_filter, bn=False):
def f(input):
if bn:
input = nn.BatchNorm2d(input.size(1), affine=False).cuda()(input)
activation = F.relu(input)
return nn.Conv2d(in_channels=input.size(1), out_channels=nb_filter, kernel_size=ns_filter, padding=1).cuda()(activation)
return f
def _residual_unit(self, nb_filter):
def f(input):
residual = self._bn_relu_conv(nb_filter=nb_filter, ns_filter=3)(input)
residual = self._bn_relu_conv(nb_filter=nb_filter, ns_filter=3)(residual)
return self._shortcut(input, residual)
return f
def ResUnits(self, residual_unit, nb_filter, repetations=1):
def f(input):
for i in range(repetations):
input = residual_unit(nb_filter=nb_filter)(input)
return input
return f
def forward(self, input):
main_inputs = []
outputs = []
c_input, p_input, t_input, e_input = input
main_inputs.append([c_input, p_input, t_input])
#########c#########
# conv1
c_conv1 = self.c_conv1(c_input)
c_residual_output = self.ResUnits(self._residual_unit, nb_filter=64, repetations=self.nb_residual_unit)(c_conv1)
# conv2
c_activation = F.relu(c_residual_output)
c_conv2 = self.c_conv2(c_activation)
outputs.append(c_conv2)
#########p#########
# conv1
p_conv1 = self.p_conv1(p_input)
p_residual_output = self.ResUnits(self._residual_unit, nb_filter=64, repetations=self.nb_residual_unit)(p_conv1)
# conv2
p_activation = F.relu(p_residual_output)
p_conv2 = self.p_conv2(p_activation)
outputs.append(p_conv2)
#########t#########
# conv1
# t_conv1 = self.t_conv1(t_input)
# t_residual_output = self.ResUnits(self._residual_unit, nb_filter=64, repetations=self.nb_residual_unit)(t_conv1)
# # conv2
# t_activation = F.relu(t_residual_output)
# t_conv2 = self.c_conv2(t_activation)
# outputs.append(t_conv2)
# parameter-matrix-based fusion
if len(outputs) == 1:
main_output = outputs[0]
else:
from iLayer import iLayer
new_outputs = []
main_output = 0
for output in outputs:
cal = iLayer((output.size(1), output.size(2), output.size(3))).cuda()(output)
new_outputs.append(cal)
main_output += cal
if self.external_dim != None and self.external_dim > 0:
# external input
main_inputs.append(e_input)
# e_output = self.relu(self.linear(e_input)).view(-1, 1, 7, 7)
# e_output = self.relu(self.deconv1(e_output))
# e_output = self.relu(self.deconv2(e_output))
# e_output = self.relu(self.deconv3(e_output))
# e_output = self.relu(self.deconv4(e_output))
# e_output = self.relu(self.deconv5(e_output))
# e_output = self.deconv6(e_output)
# main_output += e_output
# else:
# print('external_dim:', self.external_dim)
main_output = F.tanh(main_output)
return main_output
if __name__ == '__main__':
print(1)
# save_loss = []
# # c_input, p_input, t_input, e_input
# ground_truth = Variable(torch.randn(3, 2, 200, 200)).cuda()
# c_input = Variable(torch.randn(3, 4, 200, 200)).cuda()
# p_input = Variable(torch.randn(3, 4, 200, 200)).cuda()
# # t_input = Variable(torch.randn(3, 6, 200, 200)).cuda()
# e_input = Variable(torch.randn(3, 2, 200, 200)).cuda() # uncertain variable
# for i in range(1000):
# input = (c_input, p_input, None, e_input)
# main_output = stnet(input)
# optimizer.zero_grad()
# loss = criterion(main_output, ground_truth)
# loss.backward()
# optimizer.step()
# save_loss.append(loss.cpu().data.numpy())
# print(i)
# # plt.switch_backend('agg')
# plt.plot(save_loss)
# plt.show()
# plt.savefig('{}{}'.format('/mnt/data/fan/SmartST', 'loss'))