-
Notifications
You must be signed in to change notification settings - Fork 1
/
models.py
208 lines (179 loc) · 8.71 KB
/
models.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch_scatter
class Embedding_init(nn.Module):
@staticmethod
def init_emb(row, col):
w = torch.empty(row, col)
torch.nn.init.normal_(w)
w = torch.nn.functional.normalize(w)
entities_emb = nn.Parameter(w)
return entities_emb
class OverAll(nn.Module):
def __init__(self, node_size, node_hidden,
rel_size, rel_hidden,
time_size,
triple_size,
rel_matrix,
ent_matrix,
time_matrix,
dropout_rate=0, depth=2, dropout_time=0.5,
device='cpu'
):
super(OverAll, self).__init__()
self.dropout_rate = dropout_rate
self.dropout_time = dropout_time
# new adding
# rel_or_time in GraphAttention.forward
self.e_encoder = GraphAttention(node_size, rel_size, triple_size, time_size, depth=depth, device=device,
dim=node_hidden)
self.r_encoder = GraphAttention(node_size, rel_size, triple_size, time_size, depth=depth, device=device,
dim=node_hidden)
# self.t_encoder = GraphAttention(node_size, rel_size, triple_size, time_size, depth=depth, device=device,
# dim=node_hidden)
self.ent_adj = self.get_spares_matrix_by_index(ent_matrix, (node_size, node_size))
self.rel_adj = self.get_spares_matrix_by_index(rel_matrix, (node_size, rel_size))
self.time_adj = self.get_spares_matrix_by_index(time_matrix, (node_size, time_size))
self.ent_emb = self.init_emb(node_size, node_hidden)
self.rel_emb = self.init_emb(rel_size, node_hidden)
self.time_emb = self.init_emb(time_size, node_hidden)
self.try_emb = self.init_emb(1, node_hidden)
self.device = device
self.ent_adj, self.rel_adj, self.time_adj = \
map(lambda x: x.to(device), [self.ent_adj, self.rel_adj, self.time_adj])
# get prepared
@staticmethod
def get_spares_matrix_by_index(index, size):
index = torch.LongTensor(index)
adj = torch.sparse.FloatTensor(torch.transpose(index, 0, 1),
torch.ones_like(index[:, 0], dtype=torch.float), size)
# dim ??
return torch.sparse.softmax(adj, dim=1)
@staticmethod
def init_emb(*size):
entities_emb = nn.Parameter(torch.randn(size))
torch.nn.init.xavier_normal_(entities_emb)
return entities_emb
def forward(self, inputs):
# inputs = [adj_matrix, r_index, r_val, rel_matrix, ent_matrix, train_pairs]
ent_feature = torch.matmul(self.ent_adj, self.ent_emb)
rel_feature = torch.matmul(self.rel_adj, self.rel_emb)
time_feature = torch.matmul(self.time_adj, self.time_emb)
# ent_feature = torch.cat([ent_feature, rel_feature, time_feature], dim=1)
# note that time_feature and rel_feature is has the same shape of ent_feature
# the dim = node_hidden, the shape[0] = # of entities
# They are obtained by gather the linked rel/time of an entity
adj_input = inputs[0]
r_index = inputs[1]
r_val = inputs[2]
t_index = inputs[3]
opt = [self.rel_emb, adj_input, r_index, r_val]
opt2 = [self.time_emb, adj_input, t_index, r_val]
# attention opt_1 or 2
out_feature_ent = self.e_encoder([ent_feature] + opt)
out_feature_rel = self.r_encoder([rel_feature] + opt)
out_feature_time = self.e_encoder([time_feature] + opt2, 1)
# out_feature_ent2 = self.e_encoder([ent_feature] + opt)
# out_feature_rel2 = self.r_encoder([rel_feature] + opt)
# out_feature_time2 = self.t_encoder([time_feature] + opt2, 1)
# out_feature_time = self.e_encoder([time_feature] + opt)
# out_feature_time = F.dropout(out_feature_time, p=self.dropout_time, training=self.training)
from config import global_args
if global_args.dual_no_time:
out_feature_overall = out_feature_rel
else:
out_feature_overall = (out_feature_rel + out_feature_time) / 2
out_feature = torch.cat((out_feature_ent, out_feature_overall), dim=-1)
out_feature = F.dropout(out_feature, p=self.dropout_rate, training=self.training)
return out_feature
class GraphAttention(nn.Module):
def __init__(self, node_size, rel_size, triple_size, time_size,
activation=torch.tanh, use_bias=True,
attn_heads=1, dim=100,
depth=1, device='cpu'):
super(GraphAttention, self).__init__()
self.node_size = node_size
self.activation = activation
self.rel_size = rel_size
self.time_size = time_size
self.triple_size = triple_size
self.use_bias = use_bias
self.attn_heads = attn_heads
self.attn_heads_reduction = 'concat'
self.depth = depth
self.device = device
self.attn_kernels = []
node_F = dim
rel_F = dim
self.ent_F = node_F
ent_F = self.ent_F
# gate kernel Eq 9 M
self.gate_kernel = OverAll.init_emb(ent_F * (self.depth + 1), ent_F * (self.depth + 1))
self.proxy = OverAll.init_emb(64, node_F * (self.depth + 1))
if self.use_bias:
self.bias = OverAll.init_emb(1, ent_F * (self.depth + 1))
for d in range(self.depth):
self.attn_kernels.append([])
for h in range(self.attn_heads):
attn_kernel = OverAll.init_emb(node_F, 1)
self.attn_kernels[d].append(attn_kernel.to(device))
def forward(self, inputs, rel_or_time=0):
outputs = []
features = inputs[0]
rel_emb = inputs[1]
adj_index = inputs[2] # adj
index = torch.tensor(adj_index, dtype=torch.int64)
index = index.to(self.device)
# adj = torch.sparse.FloatTensor(torch.LongTensor(index),
# torch.FloatTensor(torch.ones_like(index[:,0])),
# (self.node_size, self.node_size))
sparse_indices = inputs[3] # relation index i.e. r_index
sparse_val = inputs[4] # relation value i.e. r_val
features = self.activation(features)
outputs.append(features)
for l in range(self.depth):
features_list = []
for head in range(self.attn_heads):
attention_kernel = self.attn_kernels[l][head]
####
col = self.rel_size if rel_or_time == 0 else self.time_size
rels_sum = torch.sparse.FloatTensor(
torch.transpose(torch.LongTensor(sparse_indices), 0, 1),
torch.FloatTensor(sparse_val),
(self.triple_size, col)
) # relation matrix
rels_sum = rels_sum.to(self.device)
rels_sum = torch.matmul(rels_sum, rel_emb)
neighs = features[index[:, 1]]
# selfs = features[index[:, 0]]
rels_sum = F.normalize(rels_sum, p=2, dim=1)
neighs = neighs - 2 * torch.sum(neighs * rels_sum, 1, keepdim=True) * rels_sum
# Eq.3
att1 = torch.squeeze(torch.matmul(rels_sum, attention_kernel), dim=-1)
att = torch.sparse.FloatTensor(torch.transpose(index, 0, 1), att1, (self.node_size, self.node_size))
# ??? dim ??
att = torch.sparse.softmax(att, dim=1)
# ?
# print(att1)
# print(att.data)
new_features = torch_scatter.scatter_add(
torch.transpose(neighs * torch.unsqueeze(att.coalesce().values(), dim=-1), 0, 1),
index[:, 0])
new_features = torch.transpose(new_features, 0, 1)
features_list.append(new_features)
if self.attn_heads_reduction == 'concat':
features = torch.cat(features_list)
features = self.activation(features)
outputs.append(features)
outputs = torch.cat(outputs, dim=1)
proxy_att = torch.matmul(F.normalize(outputs, dim=-1),
torch.transpose(F.normalize(self.proxy, dim=-1), 0, 1))
proxy_att = F.softmax(proxy_att, dim=-1) # eq.3
proxy_feature = outputs - torch.matmul(proxy_att, self.proxy)
if self.use_bias:
gate_rate = F.sigmoid(torch.matmul(proxy_feature, self.gate_kernel) + self.bias)
else:
gate_rate = F.sigmoid(torch.matmul(proxy_feature, self.gate_kernel))
outputs = gate_rate * outputs + (1 - gate_rate) * proxy_feature
return outputs