-
Notifications
You must be signed in to change notification settings - Fork 24
/
model.py
191 lines (149 loc) · 7.77 KB
/
model.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
from layer import dilated_inception, mixprop, CGP, graph_constructor
import torchdiffeq
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
class ODEFunc(nn.Module):
def __init__(self, stnet):
super(ODEFunc, self).__init__()
self.stnet = stnet
self.nfe = 0
def forward(self, t, x):
self.nfe += 1
x = self.stnet(x)
return x
class ODEBlock(nn.Module):
def __init__(self, odefunc, method, step_size, rtol, atol, adjoint=False, perturb=False):
super(ODEBlock, self).__init__()
self.odefunc = odefunc
self.method = method
self.step_size = step_size
self.adjoint = adjoint
self.perturb = perturb
self.atol = atol
self.rtol = rtol
def forward(self, x, t):
self.integration_time = torch.tensor([0, t]).float().type_as(x)
if self.adjoint:
out = torchdiffeq.odeint_adjoint(self.odefunc, x, self.integration_time, rtol=self.rtol, atol=self.atol,
method=self.method, options=dict(step_size=self.step_size, perturb=self.perturb))
else:
out = torchdiffeq.odeint(self.odefunc, x, self.integration_time, rtol=self.rtol, atol=self.atol,
method=self.method, options=dict(step_size=self.step_size, perturb=self.perturb))
return out[-1]
class STBlock(nn.Module):
def __init__(self, receptive_field, dilation, hidden_channels, dropout, method, time, step_size, alpha,
rtol, atol, adjoint, perturb=False):
super(STBlock, self).__init__()
self.receptive_field = receptive_field
self.intermediate_seq_len = receptive_field
self.graph = None
self.dropout = dropout
self.new_dilation = 1
self.dilation_factor = dilation
self.inception_1 = dilated_inception(hidden_channels, hidden_channels, dilation_factor=1)
self.inception_2 = dilated_inception(hidden_channels, hidden_channels, dilation_factor=1)
self.gconv_1 = CGP(hidden_channels, hidden_channels, alpha=alpha,
method=method, time=time, step_size=step_size, rtol=rtol, atol=atol,
adjoint=adjoint, perturb=perturb)
self.gconv_2 = CGP(hidden_channels, hidden_channels, alpha=alpha,
method=method, time=time, step_size=step_size, rtol=rtol, atol=atol,
adjoint=adjoint, perturb=perturb)
def forward(self, x):
x = x[..., -self.intermediate_seq_len:]
for tconv in self.inception_1.tconv:
tconv.dilation = (1, self.new_dilation)
for tconv in self.inception_2.tconv:
tconv.dilation = (1, self.new_dilation)
filter = self.inception_1(x)
filter = torch.tanh(filter)
gate = self.inception_2(x)
gate = torch.sigmoid(gate)
x = filter * gate
self.new_dilation *= self.dilation_factor
self.intermediate_seq_len = x.size(3)
x = F.dropout(x, self.dropout, training=self.training)
x = self.gconv_1(x, self.graph) + self.gconv_2(x, self.graph.transpose(1, 0))
x = nn.functional.pad(x, (self.receptive_field - x.size(3), 0))
return x
def setGraph(self, graph):
self.graph = graph
def setIntermediate(self, dilation):
self.new_dilation = dilation
self.intermediate_seq_len = self.receptive_field
class MTGODE(nn.Module):
def __init__(self, buildA_true, num_nodes, device, predefined_A=None, static_feat=None, dropout=0.3,
subgraph_size=20, node_dim=40, dilation_exponential=1, conv_channels=32, end_channels=128,
seq_length=12, in_dim=2, out_dim=12, tanhalpha=3, method_1='euler', time_1=1.2, step_size_1=0.4,
method_2='euler', time_2=1.0, step_size_2=0.25, alpha=1.0, rtol=1e-4, atol=1e-3, adjoint=False,
perturb=False, ln_affine=True):
super(MTGODE, self).__init__()
if method_1 == 'euler':
self.integration_time = time_1
self.estimated_nfe = round(self.integration_time / step_size_1)
elif method_1 == 'rk4':
self.integration_time = time_1
self.estimated_nfe = round(self.integration_time / (step_size_1 / 4.0))
else:
raise ValueError("Oops! Temporal ODE solver is invaild.")
self.buildA_true = buildA_true
self.num_nodes = num_nodes
self.dropout = dropout
self.predefined_A = predefined_A
self.seq_length = seq_length
self.ln_affine = ln_affine
self.adjoint = adjoint
self.start_conv = nn.Conv2d(in_channels=in_dim, out_channels=conv_channels, kernel_size=(1, 1))
self.gc = graph_constructor(num_nodes, subgraph_size, node_dim, device, alpha=tanhalpha, static_feat=static_feat)
self.idx = torch.arange(self.num_nodes).to(device)
max_kernel_size = 7
if dilation_exponential > 1:
self.receptive_field = int(1 + (max_kernel_size - 1) * (dilation_exponential**self.estimated_nfe - 1) / (dilation_exponential - 1))
else:
self.receptive_field = self.estimated_nfe * (max_kernel_size - 1) + 1
if ln_affine:
self.affine_weight = nn.Parameter(torch.Tensor(*(conv_channels, self.num_nodes))) # C*H
self.affine_bias = nn.Parameter(torch.Tensor(*(conv_channels, self.num_nodes))) # C*H
self.ODE = ODEBlock(ODEFunc(STBlock(receptive_field=self.receptive_field, dilation=dilation_exponential,
hidden_channels=conv_channels, dropout=self.dropout, method=method_2,
time=time_2, step_size=step_size_2, alpha=alpha, rtol=rtol, atol=atol,
adjoint=False, perturb=perturb)),
method_1, step_size_1, rtol, atol, adjoint, perturb)
self.end_conv_0 = nn.Conv2d(in_channels=conv_channels, out_channels=end_channels//2, kernel_size=(1, 1), bias=True)
self.end_conv_1 = nn.Conv2d(in_channels=end_channels//2, out_channels=end_channels, kernel_size=(1, 1), bias=True)
self.end_conv_2 = nn.Conv2d(in_channels=end_channels, out_channels=out_dim, kernel_size=(1, 1), bias=True)
if ln_affine:
self.reset_parameters()
def reset_parameters(self):
init.ones_(self.affine_weight)
init.zeros_(self.affine_bias)
def forward(self, input, idx=None):
seq_len = input.size(3)
assert seq_len == self.seq_length, 'input sequence length not equal to preset sequence length'
if self.seq_length < self.receptive_field:
input = nn.functional.pad(input, (self.receptive_field-self.seq_length, 0))
if self.buildA_true:
if idx is None:
adp = self.gc(self.idx)
else:
adp = self.gc(idx)
else:
adp = self.predefined_A
x = self.start_conv(input)
if self.adjoint:
self.ODE.odefunc.stnet.setIntermediate(dilation=1)
self.ODE.odefunc.stnet.setGraph(adp)
x = self.ODE(x, self.integration_time)
self.ODE.odefunc.stnet.setIntermediate(dilation=1)
x = x[..., -1:]
x = F.layer_norm(x, tuple(x.shape[1:]), weight=None, bias=None, eps=1e-5)
if self.ln_affine:
if idx is None:
x = torch.add(torch.mul(x, self.affine_weight[:, self.idx].unsqueeze(-1)), self.affine_bias[:, self.idx].unsqueeze(-1)) # C*H
else:
x = torch.add(torch.mul(x, self.affine_weight[:, idx].unsqueeze(-1)), self.affine_bias[:, idx].unsqueeze(-1)) # C*H
x = F.relu(self.end_conv_0(x))
x = F.relu(self.end_conv_1(x))
x = self.end_conv_2(x)
return x