-
Notifications
You must be signed in to change notification settings - Fork 3
/
model.py
200 lines (165 loc) · 7.08 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
192
193
194
195
196
197
198
199
200
import torch
from torch import nn
class Encoder(nn.Module):
r"""This class encode the raw features to the descriptors
Args:
n_ref_points (int): the number of reference points
Shape:
- x_o: :math:`[B, N, n_ref + 1]`
- ld: :math:`B`
- am: :math:`[B, L_\text{max} \times L_\text{max}, N]`
- Output: :math:`[\sum^B_i{L_i}, N, 1, 1]`
where B is the batch size, N is the size of the node feature vector, n_ref is the number of reference points
, :math:`L_i` is the length of the i-th sequence, and :math:`L_\text{max}`
is the length of the longest sequence in the batch.
"""
def __init__(self, n_ref_points):
super().__init__()
mlp1_dim = 64
self.mlp1 = nn.Sequential(
nn.Conv2d(1, mlp1_dim, (1, n_ref_points + 1)),
nn.BatchNorm2d(mlp1_dim),
nn.LeakyReLU(1e-2, inplace=True),
)
cell_size = 64
self.bilstm = nn.LSTM(mlp1_dim, cell_size, batch_first=True, bidirectional=True)
mlp2_dim = 256
self.mlp2 = nn.Sequential(
nn.Conv2d(cell_size * 2, mlp2_dim, 1),
nn.BatchNorm2d(mlp2_dim),
nn.LeakyReLU(1e-2, inplace=True),
)
gcrb_dim = [256, 512]
self.gcl = GraphConvLayer(mlp2_dim, gcrb_dim[0], activation='leaky_relu')
self.gcrb_1 = GraphConvResBlock(gcrb_dim[0], gcrb_dim[0])
self.gcrb_2 = GraphConvResBlock(gcrb_dim[0], gcrb_dim[1])
self.fc = nn.Sequential(
nn.Conv2d(gcrb_dim[-1], 512, 1),
nn.BatchNorm2d(512),
nn.LeakyReLU(1e-2, inplace=True),
nn.Dropout(0.5),
nn.Conv2d(512, 400, 1),
)
def rnn_module(self, x, ld):
x_1 = x.unsqueeze(1) # [B, 1, N, 32]
x_2 = self.mlp1(x_1) # [B, 64, N, 1]
x_2 = x_2.squeeze(3).permute(0, 2, 1) # [B, N, 64]
px = nn.utils.rnn.pack_padded_sequence(x_2, ld, enforce_sorted=False, batch_first=True)
packed_out, (ht, ct) = self.bilstm(px)
padded_out = nn.utils.rnn.pad_packed_sequence(packed_out, batch_first=True, padding_value=-float('inf'))
group_1 = []
for i, len_seq in enumerate(ld):
y = padded_out[0][i, :len_seq, :]
group_1.append(y)
x_3 = torch.cat(group_1, dim=0)
x_3 = x_3.unsqueeze(2).unsqueeze(3)
x_4 = self.mlp2(x_3) # [B, 256, 1, 1]
return x_4
def forward(self, x_o, ld, am):
x = self.rnn_module(x_o, ld)
x = self.gcl(x, ld, am)
x = self.gcrb_1(x, ld, am)
x = self.gcrb_2(x, ld, am)
group_2 = []
prev = 0
for i, len_seq in enumerate(ld):
y = x[prev:prev + len_seq, :]
g_max, _ = torch.max(y, 0)
group_2.append(g_max)
prev += len_seq
x_1 = torch.stack(group_2, 0) # [B, 512, 1, 1]
x_2 = self.fc(x_1)
x_2 = x_2.squeeze(-1).squeeze(-1)
return x_2
class GraphConvResBlock(nn.Module):
r"""This class applies a residual block containing two GraphConvLayers
Args:
input_size (int): The size of input node feature vector
output_size (int): The size of output node feature vector
Shape:
- x: :math:`[\sum^B_i{L_i}, N, 1, 1]`
- ld: :math:`B`
- am: :math:`[B, L_\text{max} \times L_\text{max}, N]`
- Output: :math:`[\sum^B_i{L_i}, N, 1, 1]`
where B is the batch size, N is the size of the node feature vector,
, :math:`L_i` is the length of the i-th sequence, and :math:`L_\text{max}`
is the length of the longest sequence in the batch.
"""
def __init__(self, input_size, output_size):
super().__init__()
self.input_size = input_size
self.output_size = output_size
self.linear = nn.Sequential(
nn.Conv2d(self.input_size, self.output_size, 1),
nn.BatchNorm2d(self.output_size)
)
self.graph_conv_layer_1 = GraphConvLayer(input_size, input_size, 0.2, 'relu')
self.graph_conv_layer_2 = GraphConvLayer(input_size, output_size)
self.activate = nn.LeakyReLU(1e-2, inplace=True)
def forward(self, x, ld, am):
# linear transformation for input
residual = self.linear(x) if self.input_size != self.output_size else x
# two GCN layers
x = self.graph_conv_layer_1(x, ld, am)
x = self.graph_conv_layer_2(x, ld, am)
x = self.activate(x + residual)
return x
class GraphConvLayer(nn.Module):
r"""This class applies a graph convolutional layer
Args:
input_size (int): The size of input node feature vector
output_size (int): The size of output node feature vector
dropout (float): a dropout layer after MLP. Default=0.0
activation (str, optional): an activation function (``'relu'`` or ``'leaky_relu'``) after MLP. Default=None.
Shape:
- x: :math:`[\sum^B_i{L_i}, N, 1, 1]`
- ld: :math:`B`
- am: :math:`[B, L_\text{max} \times L_\text{max}, N]`
- Output: :math:`[\sum^B_i{L_i}, N, 1, 1]`
where B is the batch size, N is the size of the node feature vector,
, :math:`L_i` is the length of the i-th sequence, and :math:`L_\text{max}`
is the length of the longest sequence in the batch.
"""
def __init__(self, input_size: int, output_size: int, dropout: float = 0.0, activation: str = None):
super(GraphConvLayer, self).__init__()
self.nonlinear = nn.Sequential(
nn.Conv2d(input_size, output_size, 1),
nn.BatchNorm2d(output_size)
)
if activation == 'relu':
self.nonlinear.add_module('Relu', nn.ReLU(inplace=True))
elif activation == 'leaky_relu':
self.nonlinear.add_module('Leaky_Relu', nn.LeakyReLU(inplace=True))
if dropout > 0:
self.nonlinear.add_module('Dropout', nn.Dropout(dropout))
def forward(self, x: torch.Tensor, ld: list, am: torch.Tensor) -> torch.Tensor:
x = aggregate_node(x, ld, am)
x = self.nonlinear(x)
return x
def aggregate_node(x: torch.Tensor, ld: list, am: torch.Tensor) -> torch.Tensor:
r"""
This function is used to aggregate node features in the graph.
:param x: node feature matrix
:param ld: sequence length list
:param am: adjacency matrix
:return: updated node feature matrix
Shape:
- x: :math:`[\sum^B_i{L_i}, N, 1, 1]`
- ld: :math:`B`
- am: :math:`[B, L_\text{max} \times L_\text{max}, N]`
- return: :math:`[\sum^B_i{L_i}, N, 1, 1]`
where B is the batch size, N is the size of the node feature vector,
, :math:`L_i` is the length of the i-th sequence, and :math:`L_\text{max}`
is the length of the longest sequence in the batch.
"""
x = x.squeeze()
prev, ba_group = 0, []
for j, len_seq in enumerate(ld):
y = x[prev:prev + len_seq, :] # [l, input_size]
single_am = am[j, :len_seq, :len_seq]
y = single_am @ y # [l, input_size]
ba_group.append(y)
prev += len_seq
x = torch.cat(ba_group, 0) # [B, input_size]
x = x.unsqueeze(2).unsqueeze(3) # [B, input_size, 1, 1]
return x