-
Notifications
You must be signed in to change notification settings - Fork 147
/
ATT_HGCN.py
167 lines (135 loc) · 6.39 KB
/
ATT_HGCN.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
import torch.nn as nn
import torch.nn.functional as F
import torch
import dgl
from dgl.nn.pytorch import GATConv
from . import BaseModel, register_model
from ..layers.macro_layer.SemanticConv import SemanticAttention
from ..layers.MetapathConv import MetapathConv
from ..utils.utils import extract_metapaths, get_ntypes_from_canonical_etypes
@register_model('SHGP')
class ATT_HGCN(BaseModel):
@classmethod
def build_model_from_args(cls, args, hg):
ntypes = set()
return cls(
net_schema=args.net_schema,
layer_shape=args.layer_shape,
label_keys=list(args.label.keys()),
type_fusion=args.type_fusion,
type_att_size=args.type_att_size,
)
def __init__(self, net_schema, layer_shape, label_keys, type_fusion='att', type_att_size=64):
super(ATT_HGCN, self).__init__()
self.hgc1 = HeteGCNLayer(net_schema, layer_shape[0], layer_shape[1], type_fusion, type_att_size)
self.hgc2 = HeteGCNLayer(net_schema, layer_shape[1], layer_shape[2], type_fusion, type_att_size)
self.embd2class = nn.ParameterDict()
self.bias = nn.ParameterDict()
self.label_keys = label_keys
self.layer_shape = layer_shape
for k in label_keys:
self.embd2class[k] = nn.Parameter(torch.FloatTensor(layer_shape[-2][k], layer_shape[-1][k]))
nn.init.xavier_uniform_(self.embd2class[k].data, gain=1.414)
self.bias[k] = nn.Parameter(torch.FloatTensor(1, layer_shape[-1][k]))
nn.init.xavier_uniform_(self.bias[k].data, gain=1.414)
def ini_embd2class(self):
for k in self.label_keys:
nn.init.xavier_uniform_(self.embd2class[k].data, gain=1.414)
nn.init.xavier_uniform_(self.bias[k].data, gain=1.414)
def forward(self, ft_dict, adj_dict):
attention_list = []
x_dict, attention_dict = self.hgc1(ft_dict, adj_dict)
attention_list.append((attention_dict))
x_dict = self.non_linear(x_dict)
x_dict = self.dropout_ft(x_dict, 0.5)
x_dict, attention_dict = self.hgc2(x_dict, adj_dict)
attention_list.append((attention_dict))
logits = {}
embd = {}
for k in self.label_keys:
embd[k] = x_dict[k]
logits[k] = torch.mm(x_dict[k], self.embd2class[k]) + self.bias[k]
return logits, embd, attention_list
def non_linear(self, x_dict):
y_dict = {}
for k in x_dict:
y_dict[k] = F.elu(x_dict[k])
return y_dict
def dropout_ft(self, x_dict, dropout):
y_dict = {}
for k in x_dict:
y_dict[k] = F.dropout(x_dict[k], dropout, training=self.training)
return y_dict\
class _ATT_HGCN(nn.Module):
def __init__(self):
pass
def forward(self):
pass
class HeteGCNLayer(nn.Module):
def __init__(self, net_schema, in_layer_shape, out_layer_shape, type_fusion, type_att_size):
super(HeteGCNLayer, self).__init__()
self.net_schema = net_schema
self.in_layer_shape = in_layer_shape
self.out_layer_shape = out_layer_shape
self.hete_agg = nn.ModuleDict()
for k in net_schema:
self.hete_agg[k] = HeteAggregateLayer(k, net_schema[k], in_layer_shape, out_layer_shape[k], type_fusion,
type_att_size)
def forward(self, x_dict, adj_dict):
attention_dict = {}
ret_x_dict = {}
for k in self.hete_agg.keys():
ret_x_dict[k], attention_dict[k] = self.hete_agg[k](x_dict, adj_dict[k])
return ret_x_dict, attention_dict
class HeteAggregateLayer(nn.Module):
def __init__(self, curr_k, nb_list, in_layer_shape, out_shape, type_fusion, type_att_size):
super(HeteAggregateLayer, self).__init__()
self.nb_list = nb_list
self.curr_k = curr_k
self.type_fusion = type_fusion
self.W_rel = nn.ParameterDict()
for k in nb_list:
try:
self.W_rel[k] = nn.Parameter(torch.FloatTensor(in_layer_shape[k], out_shape))
except KeyError as ke:
self.W_rel[k] = nn.Parameter(torch.FloatTensor(in_layer_shape[self.curr_k], out_shape))
finally:
nn.init.xavier_uniform_(self.W_rel[k].data, gain=1.414)
self.w_self = nn.Parameter(torch.FloatTensor(in_layer_shape[curr_k], out_shape))
nn.init.xavier_uniform_(self.w_self.data, gain=1.414)
self.bias = nn.Parameter(torch.FloatTensor(1, out_shape))
nn.init.xavier_uniform_(self.bias.data, gain=1.414)
if type_fusion == 'att':
self.w_query = nn.Parameter(torch.FloatTensor(out_shape, type_att_size))
nn.init.xavier_uniform_(self.w_query.data, gain=1.414)
self.w_keys = nn.Parameter(torch.FloatTensor(out_shape, type_att_size))
nn.init.xavier_uniform_(self.w_keys.data, gain=1.414)
self.w_att = nn.Parameter(torch.FloatTensor(2 * type_att_size, 1))
nn.init.xavier_uniform_(self.w_att.data, gain=1.414)
def forward(self, x_dict, adj_dict):
attention_curr_k = 0
self_ft = torch.mm(x_dict[self.curr_k], self.w_self)
nb_ft_list = [self_ft]
nb_name = [self.curr_k + '_self']
for k in self.nb_list:
try:
nb_ft = torch.mm(x_dict[k], self.W_rel[k])
except KeyError as ke:
nb_ft = torch.mm(x_dict[self.curr_k], self.W_rel[k])
finally:
nb_ft = torch.spmm(adj_dict[k], nb_ft)
nb_ft_list.append(nb_ft)
nb_name.append(k)
if self.type_fusion == 'mean':
agg_nb_ft = torch.cat([nb_ft.unsqueeze(1) for nb_ft in nb_ft_list], 1).mean(1)
attention = []
elif self.type_fusion == 'att':
att_query = torch.mm(self_ft, self.w_query).repeat(len(nb_ft_list), 1)
att_keys = torch.mm(torch.cat(nb_ft_list, 0), self.w_keys)
att_input = torch.cat([att_keys, att_query], 1)
att_input = F.dropout(att_input, 0.5, training=self.training)
e = F.elu(torch.matmul(att_input, self.w_att))
attention = F.softmax(e.view(len(nb_ft_list), -1).transpose(0, 1), dim=1) # 4025*3
agg_nb_ft = torch.cat([nb_ft.unsqueeze(1) for nb_ft in nb_ft_list], 1).mul(attention.unsqueeze(-1)).sum(1)
output = agg_nb_ft + self.bias
return output, attention