-
Notifications
You must be signed in to change notification settings - Fork 0
/
KGC_model.py
237 lines (184 loc) · 8.48 KB
/
KGC_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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
class KGEModel(nn.Module):
def __init__(self, sample_method,device,model_name, nentity, nrelation, hidden_dim, gamma,
double_entity_embedding=False, double_relation_embedding=False):
super(KGEModel, self).__init__()
self.model_name = model_name
self.nentity = nentity
self.nrelation = nrelation
self.hidden_dim = hidden_dim
self.epsilon = 2.0
self.device = device
self.sample_method = sample_method
self.gamma = nn.Parameter(
torch.Tensor([gamma]),
requires_grad=False
)
self.embedding_range = nn.Parameter(torch.Tensor([(self.gamma.item() + self.epsilon) / hidden_dim]),requires_grad=False)
self.entity_dim = hidden_dim * 2 if double_entity_embedding else hidden_dim
self.relation_dim = hidden_dim * 2 if double_relation_embedding else hidden_dim
self.entity_embedding = nn.Parameter(torch.zeros(nentity, self.entity_dim))
nn.init.uniform_(tensor=self.entity_embedding, a=-self.embedding_range.item(), b=self.embedding_range.item() )
self.relation_embedding = nn.Parameter(torch.zeros(nrelation, self.relation_dim))
nn.init.uniform_( tensor=self.relation_embedding, a=-self.embedding_range.item(), b=self.embedding_range.item() )
if model_name == 'pRotatE':
self.modulus = nn.Parameter(torch.Tensor([[0.5 * self.embedding_range.item()]]))
# Do not forget to modify this line when you add a new model in the "forward" function
if model_name not in ['TransE', 'DistMult', 'ComplEx', 'RotatE', 'pRotatE', 'PairRE']:
raise ValueError('model %s not supported' % model_name)
if model_name == 'RotatE' and (not double_entity_embedding or double_relation_embedding):
raise ValueError('RotatE should use --double_entity_embedding')
if model_name == 'ComplEx' and (not double_entity_embedding or not double_relation_embedding):
raise ValueError('ComplEx should use --double_entity_embedding and --double_relation_embedding')
def forward(self, sample, mode, train_or_test):
if mode == 'single':
batch_size, negative_sample_size = sample.size(0), 1
head = torch.index_select(
self.entity_embedding,
dim=0,
index=sample[:, 0]
).unsqueeze(1)
relation = torch.index_select(
self.relation_embedding,
dim=0,
index=sample[:, 1]
).unsqueeze(1)
tail = torch.index_select(
self.entity_embedding,
dim=0,
index=sample[:, 2]
).unsqueeze(1)
elif mode == 'head-batch':
# gumbel test has same code with uni.
if self.sample_method=='gumbel' and train_or_test== 'train':
tail_part, head = sample
else:
tail_part, head_part = sample
batch_size, negative_sample_size = head_part.size(0), head_part.size(1)
head = torch.index_select(
self.entity_embedding,
dim=0,
index=head_part.view(-1)
).view(batch_size, negative_sample_size, -1)
relation = torch.index_select(
self.relation_embedding,
dim=0,
index=tail_part[:, 1]
).unsqueeze(1)
tail = torch.index_select(
self.entity_embedding,
dim=0,
index=tail_part[:, 2]
).unsqueeze(1)
elif mode == 'tail-batch':
if self.sample_method=='gumbel' and train_or_test== 'train':
head_part, tail = sample
else:
head_part, tail_part = sample
batch_size, negative_sample_size = tail_part.size(0), tail_part.size(1)
tail = torch.index_select(
self.entity_embedding,
dim=0,
index=tail_part.view(-1)
).view(batch_size, negative_sample_size, -1)
head = torch.index_select(
self.entity_embedding,
dim=0,
index=head_part[:, 0]
).unsqueeze(1)
relation = torch.index_select(
self.relation_embedding,
dim=0,
index=head_part[:, 1]
).unsqueeze(1)
else:
raise ValueError('mode %s not supported' % mode)
model_func = {
'TransE': self.TransE,
'DistMult': self.DistMult,
'ComplEx': self.ComplEx,
'RotatE': self.RotatE,
'pRotatE': self.pRotatE,
'PairRE':self.PairRE,
}
if self.model_name in model_func:
score = model_func[self.model_name](head, relation, tail, mode)
else:
raise ValueError('model %s not supported' % self.model_name)
return score
def TransE(self, head, relation, tail, mode):
if mode == 'head-batch':
score = head + (relation - tail)
else:
score = (head + relation) - tail
score = self.gamma.item() - torch.norm(score, p=1, dim=2)
return score
def _transfer(self, e, e_t, r_t):
return F.normalize(e + (e * e_t).sum(dim=1, keepdim=True) * r_t, 2, -1)
def DistMult(self, head, relation, tail, mode):
if mode == 'head-batch':
score = head * (relation * tail)
else:
score = (head * relation) * tail
score = score.sum(dim=2)
return score
def ComplEx(self, head, relation, tail, mode):
re_head, im_head = torch.chunk(head, 2, dim=2)
re_relation, im_relation = torch.chunk(relation, 2, dim=2)
re_tail, im_tail = torch.chunk(tail, 2, dim=2)
if mode == 'head-batch':
re_score = re_relation * re_tail + im_relation * im_tail
im_score = re_relation * im_tail - im_relation * re_tail
score = re_head * re_score + im_head * im_score
else:
re_score = re_head * re_relation - im_head * im_relation
im_score = re_head * im_relation + im_head * re_relation
score = re_score * re_tail + im_score * im_tail
score = score.sum(dim=2)
return score
def RotatE(self, head, relation, tail, mode):
pi = 3.14159265358979323846
re_head, im_head = torch.chunk(head, 2, dim=2)
re_tail, im_tail = torch.chunk(tail, 2, dim=2)
# Make phases of relations uniformly distributed in [-pi, pi]
phase_relation = relation / (self.embedding_range.item() / pi)
re_relation = torch.cos(phase_relation)
im_relation = torch.sin(phase_relation)
if mode == 'head-batch':
re_score = re_relation * re_tail + im_relation * im_tail
im_score = re_relation * im_tail - im_relation * re_tail
re_score = re_score - re_head
im_score = im_score - im_head
else:
re_score = re_head * re_relation - im_head * im_relation
im_score = re_head * im_relation + im_head * re_relation
re_score = re_score - re_tail
im_score = im_score - im_tail
score = torch.stack([re_score, im_score], dim=0)
score = score.norm(dim=0)
score = self.gamma.item() - score.sum(dim=2)
return score
def pRotatE(self, head, relation, tail, mode):
pi = 3.14159262358979323846
# Make phases of entities and relations uniformly distributed in [-pi, pi]
phase_head = head / (self.embedding_range.item() / pi)
phase_relation = relation / (self.embedding_range.item() / pi)
phase_tail = tail / (self.embedding_range.item() / pi)
if mode == 'head-batch':
score = phase_head + (phase_relation - phase_tail)
else:
score = (phase_head + phase_relation) - phase_tail
score = torch.sin(score)
score = torch.abs(score)
score = self.gamma.item() - score.sum(dim=2) * self.modulus
return score
def PairRE(self, head, relation, tail, mode):
re_head, re_tail = torch.chunk(relation, 2, dim=2)
head = F.normalize(head, 2, -1)
tail = F.normalize(tail, 2, -1)
score = head * re_head - tail * re_tail
score = self.gamma.item() - torch.norm(score, p=1, dim=2)
return score