-
Notifications
You must be signed in to change notification settings - Fork 3
/
models.py
279 lines (218 loc) · 10.5 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
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
import torch
import torch.nn.functional as F
import torch.nn as nn
class BaseModule(nn.Module):
def __init__(self, n_ent, n_rel, args):
super(BaseModule, self).__init__()
self.lamb = args.lamb
self.p = args.p # norm
self.margin = args.margin
def init_weight(self):
for param in self.parameters():
nn.init.xavier_uniform_(param.data)
def score(self, head, tail, rela):
raise NotImplementedError
def prob_logit(self, head, tail, rela):
raise NotImplementedError
def prob(self, head, tail, rela):
return F.softmax(self.prob_logit(head, tail, rela), dim=-1)
def pair_loss(self, head, tail, rela, n_head, n_tail):
d_pos = self.score(head, tail, rela)
d_neg = self.score(n_head, n_tail, rela)
return F.relu(self.margin + d_pos - d_neg)
def softmax_loss(self, head, tail, rela):
rela = rela.unsqueeze(-1).expand_as(head)
probs = self.prob_logit(head, tail, rela)
n = probs.size(0)
truth = torch.zeros(n).type(torch.LongTensor).cuda()
truth_probs = torch.nn.LogSoftmax(-1)(probs)[torch.arange(0, n).type(torch.LongTensor).cuda(), truth]
return -truth_probs
def point_loss(self, head, tail, rela, label):
softplus = torch.nn.Softplus()
score = self.forward(head, tail, rela)
score = softplus(-1*label*score)
return score
def sigmoid_loss(self, head,tail, rela, n_head, n_tail):
logsigmoid = torch.nn.LogSigmoid()
p_score = self.forward(head, tail, rela)
n_score = self.forward(n_head, n_tail, rela.unsqueeze(1))
p_score = torch.sum(logsigmoid(p_score))
n_score = torch.sum(logsigmoid(-n_score))
return - p_score - n_score
class TransEModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(TransEModule, self).__init__(n_ent, n_rel, args)
self.rel_embed = nn.Embedding(n_rel, args.hidden_dim)
self.ent_embed = nn.Embedding(n_ent, args.hidden_dim)
self.init_weight()
def forward(self, head, tail, rela):
shape = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_embed = F.normalize(self.ent_embed(head),2,-1)
tail_embed = F.normalize(self.ent_embed(tail),2,-1)
rela_embed = self.rel_embed(rela)
return torch.norm(tail_embed - head_embed - rela_embed, p=self.p, dim=-1).view(shape)
def score(self, head, tail, rela):
return self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return -self.forward(head, tail, rela)
class TransDModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(TransDModule, self).__init__(n_ent, n_rel, args)
self.rel_embed = nn.Embedding(n_rel, args.hidden_dim)
self.ent_embed = nn.Embedding(n_ent, args.hidden_dim)
self.proj_rel_embed = nn.Embedding(n_rel, args.hidden_dim)
self.proj_ent_embed = nn.Embedding(n_ent, args.hidden_dim)
self.init_weight()
def init_weight(self):
for param in self.parameters():
nn.init.xavier_uniform_(param.data)
def _transfer(self, e, t, r):
return F.normalize(e + torch.sum(e*t, -1, True)*r, 2, -1)
def forward(self, head, tail, rela):
shape = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_e = F.normalize(self.ent_embed(head), 2, -1)
tail_e = F.normalize(self.ent_embed(tail), 2, -1)
rela_e = F.normalize(self.rel_embed(rela), 2, -1)
head_t = self.proj_ent_embed(head)
tail_t = self.proj_ent_embed(tail)
rela_t = self.proj_rel_embed(rela)
head_proj = self._transfer(head_e, head_t, rela_t)
tail_proj = self._transfer(tail_e, tail_t, rela_t)
return torch.norm(tail_proj - head_proj - rela_e, p=self.p, dim=-1).view(shape)
def score(self, head, tail, rela):
return self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return -self.forward(head, tail, rela)
class TransHModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(TransHModule, self).__init__()
self.rel_embed = nn.Embedding(n_rel, args.hidden_dim)
self.ent_embed = nn.Embedding(n_ent, args.hidden_dim)
self.proj_rel_embed = nn.Embedding(n_rel, args.hidden_dim)
self.init_weight()
def forward(self, head, tail, rela):
shape = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_embed = F.normalize(self.ent_embed(head), 2, -1)
tail_embed = F.normalize(self.ent_embed(tail), 2, -1)
rela_embed = F.normalize(self.rel_embed(rela), 2, -1)
w_embed = F.normalize(self.proj_rel_embed(rela), 2, -1)
head_proj = head_embed - torch.sum(w_embed * head_embed, dim=-1, keepdim=True) * w_embed
tail_proj = tail_embed - torch.sum(w_embed * tail_embed, dim=-1, keepdim=True) * w_embed
return torch.norm(tail_proj - head_proj - rela_embed, p=self.p, dim=-1).view(shape)
def score(self, head, tail, rela):
return self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return -self.forward(head, tail, rela)
class DistMultModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(DistMultModule, self).__init__(n_ent, n_rel, args)
self.ent_embed = nn.Embedding(n_ent, args.hidden_dim)
self.rel_embed = nn.Embedding(n_rel, args.hidden_dim)
self.init_weight()
def forward(self, head, tail, rela):
shapes = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_embed = self.ent_embed(head)
tail_embed = self.ent_embed(tail)
rela_embed = self.rel_embed(rela)
return torch.sum(tail_embed * head_embed * rela_embed, dim=-1).view(shapes)
def score(self, head, tail, rela):
return -self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return self.forward(head, tail, rela)
class ComplExModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(ComplExModule, self).__init__(n_ent, n_rel, args)
self.ent_re_embed = nn.Embedding(n_ent, args.hidden_dim)
self.ent_im_embed = nn.Embedding(n_ent, args.hidden_dim)
self.rel_re_embed = nn.Embedding(n_rel, args.hidden_dim)
self.rel_im_embed = nn.Embedding(n_rel, args.hidden_dim)
self.init_weight()
def forward(self, head, tail, rela):
shapes = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_re_embed = self.ent_re_embed(head)
tail_re_embed = self.ent_re_embed(tail)
rela_re_embed = self.rel_re_embed(rela)
head_im_embed = self.ent_im_embed(head)
tail_im_embed = self.ent_im_embed(tail)
rela_im_embed = self.rel_im_embed(rela)
score = torch.sum(rela_re_embed * head_re_embed * tail_re_embed, dim=-1) \
+ torch.sum(rela_re_embed * head_im_embed * tail_im_embed, dim=-1) \
+ torch.sum(rela_im_embed * head_re_embed * tail_im_embed, dim=-1) \
- torch.sum(rela_im_embed * head_im_embed * tail_re_embed, dim=-1)
return score.view(shapes)
def score(self, head, tail, rela):
return -self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return self.forward(head, tail, rela)
class SimplEModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(SimplEModule, self).__init__(n_ent, n_rel, args)
self.head_embed = nn.Embedding(n_ent, args.hidden_dim)
self.tail_embed = nn.Embedding(n_ent, args.hidden_dim)
self.rela_embed = nn.Embedding(n_rel, args.hidden_dim)
self.rela_inv_embed = nn.Embedding(n_rel, args.hidden_dim)
self.init_weight()
def forward(self, head, tail, rela):
shapes = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_embed = self.head_embed(head)
tail_embed = self.tail_embed(tail)
rela_embed = self.rela_embed(rela)
head_inv_embed = self.tail_embed(head)
tail_inv_embed = self.head_embed(tail)
rela_inv_embed = self.rela_inv_embed(rela)
score = torch.sum(head_embed * rela_embed * tail_embed, dim=-1) \
+ torch.sum(head_inv_embed * rela_inv_embed * tail_inv_embed, dim=-1)
return score.view(shapes)
def score(self, head, tail, rela):
return -self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return self.forward(head, tail, rela)
class RotatEModule(BaseModule):
def __init__(self, n_ent, n_rel, args):
super(RotatEModule, self).__init__(n_ent, n_rel, args)
self.ent_re_embed = nn.Embedding(n_ent, args.hidden_dim)
self.ent_im_embed = nn.Embedding(n_ent, args.hidden_dim)
self.rel_re_embed = nn.Embedding(n_rel, args.hidden_dim)
self.rel_im_embed = nn.Embedding(n_rel, args.hidden_dim)
self.init_weight()
def forward(self, head, tail, rela):
shapes = head.size()
head = head.contiguous().view(-1)
tail = tail.contiguous().view(-1)
rela = rela.contiguous().view(-1)
head_re_embed = self.ent_re_embed(head)
head_im_embed = self.ent_im_embed(head)
tail_re_embed = self.ent_re_embed(tail)
tail_im_embed = self.ent_im_embed(tail)
rela_re_embed = self.rel_re_embed(rela)
rela_im_embed = self.rel_im_embed(rela)
rela_norm = torch.sqrt(rela_re_embed**2 + rela_im_embed**2 + 1e-10)
rela_re_embed = rela_re_embed/rela_norm
rela_im_embed = rela_im_embed/rela_norm
re_score = head_re_embed*rela_re_embed - head_im_embed*rela_im_embed - tail_re_embed
im_score = head_re_embed*rela_im_embed + head_im_embed*rela_re_embed - tail_im_embed
score = torch.sqrt(re_score**2 + im_score**2 + 1e-10).sum(dim=-1)
score = self.margin - score
return score.view(shapes)
def score(self, head, tail, rela):
return -self.forward(head, tail, rela)
def prob_logit(self, head, tail, rela):
return self.forward(head, tail, rela)