-
Notifications
You must be signed in to change notification settings - Fork 8
/
models.py
230 lines (190 loc) · 8.22 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
import torch
import torch.nn as nn
import util
import numpy as np
class KGE(nn.Module):
def __init__(self,n_ents,n_rels,ent_dim,rel_dim,max_norm=False):
super(KGE,self).__init__()
self.num_ents = n_ents
if max_norm:
self.entities = nn.Embedding(n_ents,ent_dim,max_norm=1)
else:
self.entities = nn.Embedding(n_ents, ent_dim)
self.rels = nn.Embedding(n_rels,rel_dim)
self.init()
def forward(self,sources,targets,rels):
raise NotImplementedError('Abstract method')
def predict(self, batch,negs=None,is_target=True,is_pad=False):
sources,targets, rels = util.get_triples(batch,negs,
is_target=is_target,volatile=True,is_pad=is_pad)
return self.forward(sources,targets,rels)
def init(self):
self.entities.weight.data.uniform_(-0.1, 0.1)
self.rels.weight.data.uniform_(-0.1, 0.1)
def broadcast(self,sources,targets,rels):
# PyTorch 0.1.11 does not support broadcasting
rels = rels.unsqueeze(1)
if sources.size()[1] > targets.size()[1]:
rels = rels.expand_as(sources)
targets = targets.expand_as(sources)
else:
rels = rels.expand_as(targets)
sources = sources.expand_as(targets)
return sources,targets,rels
def inner_prod(self,s,r,t):
r = r.unsqueeze(1).expand_as(t)
prod = torch.mul(r,t).transpose(1,2)
score = torch.bmm(s, prod)
return score
def all_entity_vectors(self):
var = util.to_var(range(self.num_ents),volatile=True)
entities = self.entities(var).data.cpu().numpy()
return entities
def entity_vectors(self,ids):
var = util.to_var(ids,volatile=True)
vector = self.entities(var).data.cpu().numpy()
return vector
def relation_vectors(self,ids):
var = util.to_var(ids,volatile=True)
vector = self.rels(var).data.cpu().numpy()
return vector
class Rescal(KGE):
def __init__(self,n_ents,n_rels,ent_dim):
super(Rescal,self).__init__(n_ents,n_rels,ent_dim,ent_dim*ent_dim,max_norm=True)
self.dim = ent_dim
print("Initializing RESCAL model")
def init(self):
#self.entities.weight.data.normal_(std=0.1)
#self.rels.weight.data.normal_(std=0.1)
self.entities.weight.data.uniform_(-1., 1.)
self.rels.weight.data.uniform_(-1., 1.)
def forward(self,sources,targets,rels):
sources = self.entities(sources)
targets = self.entities(targets)
rels = self.rels(rels)
#Reshape rels
rels = rels.view(-1,self.dim,self.dim)
#score = x_s^T W_r x_t
out = torch.bmm(torch.bmm(sources,rels),torch.transpose(targets,1,2))
# First element is positive, rest are negatives
out = out.view(-1,out.size()[1]*out.size()[2])
return out
def output(self,entities,rels,is_target):
'''
Given source and rels output the target or given target and rels output the source vector
:param entities: source or target entity ids
:param rels: rel ids
:param is_target: True for predicting targets
:return:
'''
entities = self.entities(util.to_var(entities,True)).unsqueeze(2)
rels = self.rels(util.to_var(rels,True))
# Reshape rels
rels = rels.view(-1, self.dim, self.dim)
if is_target:
out = torch.bmm(torch.transpose(entities, 1, 2),rels)
else:
out = torch.bmm(rels, entities)
out = out.view(-1, out.size()[1] * out.size()[2])
return out.data.cpu().numpy()
class TransE(KGE):
def __init__(self, n_ents, n_rels, ent_dim):
super(TransE, self).__init__(n_ents, n_rels, ent_dim, ent_dim,max_norm=True)
print("Initializing TransE model")
def forward(self,sources,targets,rels):
sources = self.entities(sources)
targets = self.entities(targets)
rels = self.rels(rels)
sources,targets,rels = self.broadcast(sources,targets,rels)
# score = -||x_s + x_r - x_t||_2
d = sources + rels - targets
#d = torch.abs(d)
d = torch.mul(d, d)
d = torch.sum(d,2)
return torch.neg(d)
def output(self,entities,rels,is_target):
entities = self.entities(util.to_var(entities, True)).unsqueeze(2)
rels = self.rels(util.to_var(rels, True))
if is_target:
out = entities + rels
else:
out = entities - rels
return out.view(-1,out.size()[1] * out.size()[2]).data.cpu().numpy()
class Distmult(KGE):
def __init__(self, n_ents, n_rels, ent_dim):
super(Distmult, self).__init__(n_ents, n_rels, ent_dim, ent_dim,max_norm=True)
print("Initializing Distmult model")
def forward(self, sources, targets, rels):
sources = self.entities(sources)
targets = self.entities(targets)
rels = self.rels(rels)
# score = x_s^T Diag(W_r) x_t
if sources.size()[1] > targets.size()[1]:
out = self.inner_prod(sources,rels,targets).squeeze(2)
else:
out = self.inner_prod(targets, rels, sources).squeeze(2)
return out
def output(self,entities,rels,is_target):
entities = self.entities(util.to_var(entities, True)).unsqueeze(2)
rels = self.rels(util.to_var(rels, True))
out = torch.mul(entities,rels)
return out.squeeze(2).data.cpu().numpy()
def init(self):
#self.entities.weight.data.normal_(std=0.1)
#self.rels.weight.data.normal_(std=0.1)
self.entities.weight.data.normal_()
self.rels.weight.data.normal_()
class ComplEx(KGE):
def __init__(self, n_ents, n_rels, ent_dim):
super(ComplEx, self).__init__(n_ents, n_rels, ent_dim, ent_dim,max_norm=True)
self.entities_i = nn.Embedding(n_ents,ent_dim,max_norm=1)
self.rels_i = nn.Embedding(n_rels,ent_dim)
self.init_all()
print("Initializing ComplEx model")
def init_all(self):
self.entities.weight.data.normal_()
self.rels.weight.data.normal_()
self.entities_i.weight.data.normal_()
self.rels_i.weight.data.normal_()
def forward(self,sources,targets,rels):
# score = <w_r,e_s,e_o> + <w_r,e_s_i,e_o_i> + <w_r_i,e_s,e_o_i> - <w_r_i,e_s_i,e_o>
if sources.size()[1] > targets.size()[1]:
return self.complex(targets, rels, sources)
return self.complex(sources, rels, targets)
def complex(self, sources, rels, targets):
sources_i = self.entities_i(sources)
targets_i = self.entities_i(targets)
rels_i = self.rels_i(rels)
sources = self.entities(sources)
targets = self.entities(targets)
rels = self.rels(rels)
out = self.inner_prod(targets, rels, sources) \
+ self.inner_prod(targets_i, rels, sources_i) \
+ self.inner_prod(targets_i, rels_i, sources) \
- self.inner_prod(targets, rels_i, sources_i)
out = out.squeeze(2)
return out
def output(self,entities,rels,is_target):
entities_i = self.entities_i(util.to_var(entities, True)).unsqueeze(2)
rels_i = self.rels_i(util.to_var(rels, True))
entities = self.entities(util.to_var(entities, True)).unsqueeze(2)
rels = self.rels(util.to_var(rels, True))
if is_target:
out = torch.mul(entities,rels) + torch.mul(entities_i,rels) \
+ torch.mul(rels_i,entities) - torch.mul(rels_i,entities_i)
else:
out = torch.mul(entities, rels) + torch.mul(entities_i, rels) \
+ torch.mul(rels_i, entities_i) - torch.mul(rels_i, entities)
return out.squeeze(2).data.cpu().numpy()
def entity_vectors(self, ids):
var = util.to_var(ids, volatile=True)
ents_r = self.entities(var).data.cpu().numpy()
ents_i = self.entities_i(var).data.cpu().numpy()
ents_v = np.concatenate((ents_r, ents_i), axis=1)
return ents_v
def relation_vectors(self,ids):
var = util.to_var(ids,volatile=True)
rels_r = self.rels(var).data.cpu().numpy()
rels_i = self.rels_i(var).data.cpu().numpy()
rels_v = np.concatenate((rels_r,rels_i),axis=1)
return rels_v