-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
285 lines (251 loc) · 10.5 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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import BatchNorm
from torch_geometric.nn.conv import MessagePassing
from torch_geometric.utils import softmax
from torch_geometric.nn.inits import glorot, zeros
def mask(x, mask_rate=0.5):
num_nodes = x.size(0)
perm = torch.randperm(num_nodes, device=x.device)
num_mask_nodes = int(mask_rate * num_nodes)
mask_nodes = perm[: num_mask_nodes]
return mask_nodes
def sce_loss(x, y, alpha=1):
x = F.normalize(x, p=2, dim=-1)
y = F.normalize(y, p=2, dim=-1)
loss = (1 - (x * y).sum(dim=-1)).pow_(alpha)
# loss = -(x * y).sum(dim=-1)
loss = loss.mean()
return loss
class GAT(MessagePassing):
def __init__(self, in_channels, dropout, bias=True, **kwargs):
kwargs.setdefault('aggr', 'add')
super().__init__(node_dim=0, **kwargs)
self.in_channels = in_channels
self.dropout = dropout
self.att_src = nn.Parameter(torch.Tensor(1, in_channels))
self.att_dst = nn.Parameter(torch.Tensor(1, in_channels))
if bias:
self.bias = nn.Parameter(torch.Tensor(in_channels))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self):
glorot(self.att_src)
glorot(self.att_dst)
zeros(self.bias)
def forward(self, x, edge_index):
x_src, x_dst = x
alpha_src = (x_src * self.att_src).sum(-1)
alpha_dst = (x_dst * self.att_dst).sum(-1)
alpha = (alpha_src, alpha_dst)
out = self.propagate(edge_index, x=x, alpha=alpha, size=None)
if self.bias is not None:
out += self.bias
return out
def message(self, x_j, alpha_j, alpha_i, index, ptr, size_i):
alpha = alpha_j + alpha_i
alpha = F.leaky_relu(alpha)
alpha = softmax(alpha, index, ptr, size_i)
alpha = F.dropout(alpha, p=self.dropout, training=self.training)
return x_j * alpha.unsqueeze(-1)
class Cross_View(nn.Module):
def __init__(self, data, hidden_dim, feat_drop, att_drop1, att_drop2, r1, r2, r3):
super(Cross_View, self).__init__()
self.data = data
self.hidden_dim = hidden_dim
self.r1 = r1
self.r2 = r2
self.r3 = r3
self.fc = nn.ModuleDict({
n_type: nn.Linear(
data[n_type].x.shape[1],
hidden_dim,
bias=True
)
for n_type in data.use_nodes
})
self.feat_drop = nn.Dropout(feat_drop)
self.enc_mask_token = nn.Parameter(torch.zeros(1, hidden_dim))
# mask gnn
self.intra = nn.ModuleList([
GAT(hidden_dim, att_drop1)
for _ in range(len(self.data.schema_dict))
])
self.action = nn.ModuleList([
nn.PReLU() for _ in range(len(self.data.schema_dict))
])
self.act = nn.ModuleDict({
s: nn.PReLU() for s in self.data.use_nodes
})
self.bn = nn.ModuleList([
BatchNorm(hidden_dim) for _ in range(len(self.data.schema_dict))
])
self.schema_dict = {s: i for i, s in enumerate(self.data.schema_dict)}
self.reset_parameter()
# message passing gnn
self.intra_mp = nn.ModuleList([
GAT(hidden_dim, att_drop2)
for _ in range(len(self.data.mp))
])
self.action_mp = nn.ModuleList([
nn.PReLU() for _ in range(len(self.data.mp))
])
self.bn_mp = nn.ModuleList([
BatchNorm(hidden_dim) for _ in range(len(self.data.mp))
])
self.mp = {s: i for i, s in enumerate(self.data.mp)}
def reset_parameter(self):
for fc in self.fc.values():
nn.init.xavier_normal_(fc.weight, gain=1.414)
def forward(self, data):
h = {}
for n_type in data.use_nodes:
h[n_type] = self.act[n_type](
self.feat_drop(
self.fc[n_type](data[n_type].x)
)
)
# Reserve Target Node Information
for n_type in data.mp:
src, dst = n_type
x = h[src], h[dst]
embed1 = self.intra_mp[self.mp[n_type]](x, data[n_type].edge_index)
embed1 = self.bn_mp[self.mp[n_type]](embed1)
h[dst] = self.action_mp[self.mp[n_type]](embed1)
#########################################################################
# Reconstruct
# mask_node = mask(h[data.main_node], mask_rate=self.r1)
# main_h = h[data.main_node].clone()
# main_h[mask_node] = 0.0
# main_h[mask_node] += self.enc_mask_token
#
# # sc = ('a', 'p')
# sc = ('actor', 'movie')
# src, dst = sc
# x = h[src], main_h
# # edge_index, edge_mask = dropout_edge(data[sc].edge_index, 0.1)
# embed1 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
# embed1 = self.bn[self.schema_dict[sc]](embed1)
# embed1 = self.action[self.schema_dict[sc]](embed1)
#
# # sc = ('s', 'p')
# sc = ('director', 'movie')
# src, dst = sc
# x = h[src], h[dst]
# embed2 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
# embed2 = self.bn[self.schema_dict[sc]](embed2)
# embed2 = self.action[self.schema_dict[sc]](embed2)
# loss1 = sce_loss(embed1[mask_node], embed2[mask_node].detach())
# ###########################################################################
# mask_node = mask(h[data.main_node], mask_rate=self.r2)
# main_h = h[data.main_node].clone()
# main_h[mask_node] = 0.0
# main_h[mask_node] += self.enc_mask_token
#
# # sc = ('s', 'p')
# sc = ('director', 'movie')
# src, dst = sc
# x = h[src], main_h
# # edge_index, edge_mask = dropout_edge(data[sc].edge_index, 0.05)
# embed1 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
# embed1 = self.bn[self.schema_dict[sc]](embed1)
# embed1 = self.action[self.schema_dict[sc]](embed1)
#
# # sc = ('a', 'p')
# sc = ('actor', 'movie')
# src, dst = sc
# x = h[src], h[dst]
# embed2 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
# embed2 = self.bn[self.schema_dict[sc]](embed2)
# embed2 = self.action[self.schema_dict[sc]](embed2)
# loss2 = sce_loss(embed1[mask_node], embed2[mask_node].detach())
# return loss1 + loss2
########################################################################
# Aminer/DBLP
mask_node = mask(h[data.main_node], mask_rate=self.r1)
main_h = h[data.main_node].clone()
main_h[mask_node] = 0.0
h1 = 0
for n_type in data.schema_dict1:
src, dst = n_type
x = h[src], h[dst]
embed1 = self.intra[self.schema_dict[n_type]](x, data[n_type].edge_index)
embed1 = self.bn[self.schema_dict[n_type]](embed1)
h1 += self.action[self.schema_dict[n_type]](embed1)
sc = ('C', 'P')
src, dst = sc
x = h[src], main_h
embed2 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
embed2 = self.bn[self.schema_dict[sc]](embed2)
embed2 = self.action[self.schema_dict[sc]](embed2)
loss1 = sce_loss(embed2[mask_node], h1[mask_node].detach())
##########################################################################
mask_node = mask(h[data.main_node], mask_rate=self.r2)
main_h = h[data.main_node].clone()
main_h[mask_node] = 0.0
h1 = 0
for n_type in data.schema_dict2:
src, dst = n_type
x = h[src], h[dst]
embed1 = self.intra[self.schema_dict[n_type]](x, data[n_type].edge_index)
embed1 = self.bn[self.schema_dict[n_type]](embed1)
h1 += self.action[self.schema_dict[n_type]](embed1)
sc = ('R', 'P')
src, dst = sc
x = h[src], main_h
embed2 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
embed2 = self.bn[self.schema_dict[sc]](embed2)
embed2 = self.action[self.schema_dict[sc]](embed2)
loss2 = sce_loss(embed2[mask_node], h1[mask_node].detach())
##########################################################################
mask_node = mask(h[data.main_node], mask_rate=self.r3)
main_h = h[data.main_node].clone()
main_h[mask_node] = 0.0
h1 = 0
for n_type in data.schema_dict3:
src, dst = n_type
x = h[src], h[dst]
embed1 = self.intra[self.schema_dict[n_type]](x, data[n_type].edge_index)
embed1 = self.bn[self.schema_dict[n_type]](embed1)
h1 += self.action[self.schema_dict[n_type]](embed1)
sc = ('A', 'P')
src, dst = sc
x = h[src], main_h
embed2 = self.intra[self.schema_dict[sc]](x, data[sc].edge_index)
embed2 = self.bn[self.schema_dict[sc]](embed2)
embed2 = self.action[self.schema_dict[sc]](embed2)
loss3 = sce_loss(embed2[mask_node], h1[mask_node].detach())
return loss1 + loss2 + loss3
def get_embed(self, data):
h = {}
for n_type in data.use_nodes:
h[n_type] = self.act[n_type](
self.feat_drop(
self.fc[n_type](data[n_type].x)
)
)
h2 = 0.0
for n_type in data.mp:
src, dst = n_type
x = h[src], h[dst]
embed1 = self.intra_mp[self.mp[n_type]](x, data[n_type].edge_index)
embed1 = self.bn_mp[self.mp[n_type]](embed1)
h[dst] = self.action_mp[self.mp[n_type]](embed1)
for n_type in data.schema_dict:
src, dst = n_type
x = h[src], h[dst]
embed1 = self.intra[self.schema_dict[n_type]](x, data[n_type].edge_index)
embed1 = self.bn[self.schema_dict[n_type]](embed1)
h2 += self.action[self.schema_dict[n_type]](embed1)
return h2.detach()
def get_C(self, data):
h = {}
for n_type in data.use_nodes:
h[n_type] = self.act[n_type](
self.feat_drop(
self.fc[n_type](data[n_type].x)
)
)
return h['C'].detach()