-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathbaselines.py
140 lines (103 loc) · 4.86 KB
/
baselines.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
import torch
import torch.nn as nn
import torch.nn.functional as F
def _L2_loss_mean(x):
return torch.mean(torch.sum(torch.pow(x, 2), dim=1, keepdim=False) / 2.)
class TransE(nn.Module):
def __init__(self, args, n_entities, n_relations, device=None):
super(TransE, self).__init__()
self.device = device
self.n_entities = n_entities
self.n_relations = n_relations
self.embed_dim = args.embed_dim
self.relation_dim = args.relation_dim
self.kg_l2loss_lambda = args.kg_l2loss_lambda
self.training_neg_rate = args.training_neg_rate
self.entity_embed = nn.Embedding(
self.n_entities, self.embed_dim)
self.relation_embed = nn.Embedding(self.n_relations, self.relation_dim)
nn.init.xavier_uniform_(self.entity_embed.weight)
nn.init.xavier_uniform_(self.relation_embed.weight)
def calc_triplet_loss(self, h, r, pos_t, neg_t):
"""
h: (kg_batch_size)
r: (kg_batch_size)
pos_t: (kg_batch_size)
neg_t: (kg_batch_size)
"""
r_embed = self.relation_embed(r) # (kg_batch_size, relation_dim)
all_embed = self.entity_embed.weight
head_embed = all_embed[h] # (batch_size, concat_dim)
tail_pos_embed = all_embed[pos_t] # (batch_size, concat_dim)
tail_neg_embed = all_embed[neg_t] # (batch_size, concat_dim)
# Trans E
pos_score = torch.sum(
torch.pow(head_embed + r_embed - tail_pos_embed, 2), dim=1) # (kg_batch_size)
neg_score = torch.sum(
torch.pow(head_embed + r_embed - tail_neg_embed, 2), dim=1) # (kg_batch_size)
# triplet_loss = F.softplus(pos_score - neg_score)
triplet_loss = (-1.0) * F.logsigmoid(neg_score - pos_score)
triplet_loss = torch.mean(triplet_loss)
l2_loss = _L2_loss_mean(head_embed) + _L2_loss_mean(r_embed) + _L2_loss_mean(tail_pos_embed) + _L2_loss_mean(
tail_neg_embed)
loss = triplet_loss + self.kg_l2loss_lambda * l2_loss
return loss
def forward(self, *input):
return self.calc_triplet_loss(*input)
class TransR(nn.Module):
def __init__(self, args, n_entities, n_relations, device=None):
super(TransR, self).__init__()
self.device = device
self.n_entities = n_entities
self.n_relations = n_relations
self.embed_dim = args.embed_dim
self.relation_dim = args.relation_dim
self.kg_l2loss_lambda = args.kg_l2loss_lambda
self.training_neg_rate = args.training_neg_rate
self.entity_embed = nn.Embedding(
self.n_entities, self.embed_dim)
self.relation_embed = nn.Embedding(self.n_relations, self.relation_dim)
# self.trans_M = nn.Parameter(torch.Tensor(
# self.n_relations, self.embed_dim, self.relation_dim))
self.trans_M = nn.Parameter(torch.Tensor(
self.n_relations, self.embed_dim,
self.relation_dim))
nn.init.xavier_uniform_(self.entity_embed.weight)
nn.init.xavier_uniform_(self.relation_embed.weight)
# nn.init.xavier_uniform_(self.trans_M)
nn.init.xavier_uniform_(self.trans_M)
def calc_triplet_loss(self, h, r, pos_t, neg_t):
"""
h: (kg_batch_size)
r: (kg_batch_size)
pos_t: (kg_batch_size)
neg_t: (kg_batch_size)
"""
r_embed = self.relation_embed(r) # (kg_batch_size, relation_dim)
W_r = self.trans_M[r] # (kg_batch_size, embed_dim, relation_dim)
all_embed = self.entity_embed.weight
head_embed = all_embed[h] # (batch_size, concat_dim)
tail_pos_embed = all_embed[pos_t] # (batch_size, concat_dim)
tail_neg_embed = all_embed[neg_t] # (batch_size, concat_dim)
r_mul_h = torch.bmm(head_embed.unsqueeze(1), W_r).squeeze(
1) # (kg_batch_size, relation_dim)
r_mul_pos_t = torch.bmm(tail_pos_embed.unsqueeze(1), W_r).squeeze(
1) # (kg_batch_size, relation_dim)
r_mul_neg_t = torch.bmm(tail_neg_embed.unsqueeze(1), W_r).squeeze(
1) # (kg_batch_size, relation_dim)
# Trans R
# Equation (1)
pos_score = torch.sum(
torch.pow(r_mul_h + r_embed - r_mul_pos_t, 2), dim=1) # (kg_batch_size)
neg_score = torch.sum(
torch.pow(r_mul_h + r_embed - r_mul_neg_t, 2), dim=1) # (kg_batch_size)
# Equation (2)
# triplet_loss = F.softplus(pos_score - neg_score)
triplet_loss = (-1.0) * F.logsigmoid(neg_score - pos_score)
triplet_loss = torch.mean(triplet_loss)
l2_loss = _L2_loss_mean(r_mul_h) + _L2_loss_mean(r_embed) + _L2_loss_mean(r_mul_pos_t) + _L2_loss_mean(
r_mul_neg_t)
loss = triplet_loss + self.kg_l2loss_lambda * l2_loss
return loss
def forward(self, *input):
return self.calc_triplet_loss(*input)