-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathjoint_model.py
207 lines (186 loc) · 11.9 KB
/
joint_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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Aug 20 12:42:33 2018
@author: xuweijia
"""
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from utils import StableBCELoss,to_var
# not use self.cuda !!!!
# sampler (iterare class) no ()
# stored int key (json) --> str
class MS_model(nn.Module):
def __init__(self,args,word_dict,id2f,vocab_size,entity_size,r_size,n_neg):
super(MS_model, self).__init__()
self.args=args
self.V=vocab_size
self.E=entity_size
self.R=r_size
self.I=self.V+self.E
self.freq=id2f
self.h=args.embedding_size
self.p=np.array(list(self.freq.values()))**(3/4)
self.mysampler=torch.utils.data.sampler.WeightedRandomSampler(self.p,n_neg,replacement=False) # sample from 0-len(weight)
self.mode=args.train_mode
self.embedding=nn.Embedding(self.I,self.h,padding_idx=0)
self.r_embedding=nn.Embedding(self.R,self.h)
self.context_embedding=nn.Embedding(self.I,self.h,padding_idx=0) # context words embedding
self.use_cuda=args.use_cuda
self.bceloss=StableBCELoss()
embedding = self.embedding.weight.data
if args.using_embed:
name=args.name
n_in_embed=0
if name=='word2vec_glove':
skip_flag=True
elif name=='glove':
skip_flag=False
print(50*'*'+'loading embedding'+50*'*')
with open(args.embedding_file) as f:
n=0
for line in f:
parsed = line.rstrip().split(' ')
if n==0 and skip_flag:
n+=1
continue
assert(len(parsed) == self.h + 1)
# w in embed_file
w = parsed[0]
if w in word_dict:
n_in_embed+=1
vec=parsed[1:]
vec = torch.Tensor([float(i) for i in vec])
embedding[word_dict[w]].copy_(vec)
n+=1
print('words in embeddings / all words :{}/{}'.format(n_in_embed,self.V))
#def forward(self,centers,contexts,KB_batch,Q_centers,Q_contexts,name_KB_batch,n_neg=10,margin=7):
def forward(self,inputs):
n_neg=self.args.n_neg
margin=self.args.margin if self.mode=='ptranse' else self.args.doc_margin
loss=0
# 1 KB loss
if inputs.get('KB_batch')!=None:
KB_batch=inputs.get('KB_batch')
KB_batch=np.array(KB_batch)
pos_h=KB_batch[:,0] # (B,3)
pos_r=KB_batch[:,1]
pos_t=KB_batch[:,2]
pos_h=to_var(pos_h,self.use_cuda)
pos_r=to_var(pos_r,self.use_cuda)
pos_t=to_var(pos_t,self.use_cuda)
pos_h_embedding=self.embedding(pos_h) # h: B,h
pos_r_embedding=self.r_embedding(pos_r) # r: B,h
pos_t_embedding=self.embedding(pos_t) # t: B,h
neg_h=np.random.choice(list(range(self.V,self.V+self.E)),size=n_neg,replace=False)
neg_h_embedding=self.embedding(to_var(neg_h,self.use_cuda)) # neg_h: n,h
neg_t=np.random.choice(list(range(self.V,self.V+self.E)),size=n_neg,replace=False)
neg_t_embedding=self.embedding(to_var(neg_t,self.use_cuda)) # neg_t: n,h
neg_r=np.random.choice(self.R,size=n_neg,replace=True)
neg_r_embedding=self.r_embedding(to_var(neg_r,self.use_cuda)) # neg_r: n,h
pos_score=margin-0.5*torch.sum((pos_h_embedding+pos_r_embedding-pos_t_embedding)**2,1) # B,1 z=b-0.5|h+r-t|**2 P=σ(z(h,r,t))
# h'+r-t B,n_neg,h --> B,n_neg
neg_h_score=margin-0.5*torch.sum((neg_h_embedding.unsqueeze(0).expand(len(pos_h),n_neg,self.h) + (pos_r_embedding-pos_t_embedding).unsqueeze(1).expand(len(pos_h),n_neg,self.h))**2,-1).squeeze(-1)
neg_t_score=margin-0.5*torch.sum(((pos_h_embedding+pos_r_embedding).unsqueeze(1).expand(len(pos_h),n_neg,self.h)-neg_t_embedding.unsqueeze(0).expand(len(pos_h),n_neg,self.h))**2,-1).squeeze(-1)
neg_r_score=margin-0.5*torch.sum(((pos_h_embedding-pos_t_embedding).unsqueeze(1).expand(len(pos_h),n_neg,self.h)+neg_r_embedding.unsqueeze(0).expand(len(pos_h),n_neg,self.h))**2,-1).squeeze(-1)
# loss
target=to_var(torch.ones(pos_score.size()).long(),self.use_cuda).float()
target_neg=to_var(torch.zeros(neg_h_score.size()).long(),self.use_cuda).float()
loss_KB=self.bceloss(pos_score,target)*3 +(self.bceloss(neg_h_score,target_neg)+self.bceloss(neg_t_score,target_neg)+self.bceloss(neg_r_score,target_neg)) * n_neg
# loss_KB=self.bceloss(pos_score,target)*3 +(self.bceloss(neg_h_score,target_neg)+self.bceloss(neg_t_score,target_neg)+self.bceloss(neg_r_score,target_neg))
loss+=loss_KB
if self.mode=='ptranse':
return loss_KB
# 2 text loss
if inputs.get('context_batch')!=None:
centers,contexts=inputs.get('context_batch')
centers=to_var(centers,self.use_cuda)
contexts=to_var(contexts,self.use_cuda)
ceter_embedding=self.embedding(centers) # B,h
context_embedding=self.context_embedding(contexts) # B,h context embedding
# negative index
neg_index=[]
for i in self.mysampler:
neg_index.append(i)
neg_index=to_var(neg_index,self.use_cuda) # n neg_samples index
neg_embedding=self.context_embedding(neg_index) # n,h context embedding
# pos score
pos_score=margin-0.5*torch.sum((ceter_embedding-context_embedding)**2,1) # B,1 z=b-0.5|v-w|**2 P=σ(z(w , v))
# neg score
center_expand=ceter_embedding.unsqueeze(1).expand(len(centers),n_neg,self.h) # B,n,h each n,h, all line is this ex's vec
neg_expand=neg_embedding.expand_as(center_expand) # B,n,h all batches are same, ex's n neg vecs
neg_score=margin-0.5*torch.sum((center_expand-neg_expand)**2,-1).squeeze(-1) # B,n each line: ex to n_neg score : z_neg1 z_neg2 ... z_neg_nneg
target=to_var(torch.ones(pos_score.size()).long(),self.use_cuda).float()
target_neg=to_var(torch.zeros(neg_score.size()).long(),self.use_cuda).float()
# yi * log sigmoid(z) + (1-yi) * log (1-sigmoid(z))
loss_text=self.bceloss(pos_score,target) + self.bceloss(neg_score,target_neg)*n_neg # torch.mean(torch.log(1/(torch.exp(neg_score)+1))) p(0)=1/(1+ex) *n_neg?
# loss_text=self.bceloss(pos_score,target) + self.bceloss(neg_score,target_neg)
loss+=loss_text
# 3 Q_text_loss
if inputs.get('Q_context_batch')!=None:
Q_centers,Q_contexts=inputs.get('Q_context_batch')
#if self.mode=='joint' or self.mode=='just_anchor':
centers=to_var(Q_centers,self.use_cuda)
contexts=to_var(Q_contexts,self.use_cuda)
ceter_embedding=self.embedding(centers) # B,h
context_embedding=self.context_embedding(contexts) # B,h context embedding
# negative index (just from V)
neg_index=[]
for i in self.mysampler:
neg_index.append(i)
neg_index=to_var(neg_index,self.use_cuda) # n neg_samples index
neg_embedding=self.context_embedding(neg_index) # n,h context embedding
# pos score
pos_score=margin-0.5*torch.sum((ceter_embedding-context_embedding)**2,1) # B,1 z=b-0.5|v-w|**2 P=σ(z(w , v))
# neg score
center_expand=ceter_embedding.unsqueeze(1).expand(len(centers),n_neg,self.h) # B,n,h each n,h, all line is this ex's vec
neg_expand=neg_embedding.expand_as(center_expand) # B,n,h all batches are same, ex's n neg vecs
neg_score=margin-0.5*torch.sum((center_expand-neg_expand)**2,-1).squeeze(-1) # B,n each line: ex to n_neg score : z_neg1 z_neg2 ... z_neg_nneg
# loss yi * log sigmoid(z) + (1-yi) * log (1-sigmoid(z))
target=to_var(torch.ones(pos_score.size()).long(),self.use_cuda).float()
target_neg=to_var(torch.zeros(neg_score.size()).long(),self.use_cuda).float()
loss_Qtext=self.bceloss(pos_score,target) + self.bceloss(neg_score,target_neg) * n_neg
# loss_Qtext=self.bceloss(pos_score,target) + self.bceloss(neg_score,target_neg)
loss+=loss_Qtext
# 4 name KB loss
# if self.mode=='joint' or self.mode=='just_name_KB':
if inputs.get('KB_name_batch')!=None:
name_KB_batch=inputs.get('KB_name_batch')
name_KB_batch=np.array(name_KB_batch)
pos_h=name_KB_batch[:,0] # (B,3)
pos_r=name_KB_batch[:,1]
pos_t=name_KB_batch[:,2]
pos_h=to_var(pos_h,self.use_cuda)
pos_r=to_var(pos_r,self.use_cuda)
pos_t=to_var(pos_t,self.use_cuda)
pos_h_embedding=self.embedding(pos_h) # h: B,h
pos_r_embedding=self.r_embedding(pos_r) # r: B,h
pos_t_embedding=self.embedding(pos_t) # t: B,h
neg_h=np.random.choice(self.I,size=n_neg,replace=False)
neg_h_embedding=self.embedding(to_var(neg_h,self.use_cuda)) # neg_h: n,h
neg_t=np.random.choice(self.I,size=n_neg,replace=False)
neg_t_embedding=self.embedding(to_var(neg_t,self.use_cuda)) # neg_t: n,h
neg_r=np.random.choice(self.R,size=n_neg,replace=True)
neg_r_embedding=self.r_embedding(to_var(neg_r,self.use_cuda)) # neg_r: n,h
pos_score=margin-0.5*torch.sum((pos_h_embedding+pos_r_embedding-pos_t_embedding)**2,1) # B,1 z=b-0.5|h+r-t|**2 P=σ(z(h,r,t))
# h'+r-t B,n_neg,h --> B,n_neg
neg_h_score=margin-0.5*torch.sum((neg_h_embedding.unsqueeze(0).expand(len(pos_h),n_neg,self.h) + (pos_r_embedding-pos_t_embedding).unsqueeze(1).expand(len(pos_h),n_neg,self.h))**2,-1).squeeze(-1)
neg_t_score=margin-0.5*torch.sum(((pos_h_embedding+pos_r_embedding).unsqueeze(1).expand(len(pos_h),n_neg,self.h)-neg_t_embedding.unsqueeze(0).expand(len(pos_h),n_neg,self.h))**2,-1).squeeze(-1)
neg_r_score=margin-0.5*torch.sum(((pos_h_embedding-pos_t_embedding).unsqueeze(1).expand(len(pos_h),n_neg,self.h)+neg_r_embedding.unsqueeze(0).expand(len(pos_h),n_neg,self.h))**2,-1).squeeze(-1)
# loss
target=to_var(torch.ones(pos_score.size()).long(),self.use_cuda).float()
target_neg=to_var(torch.zeros(neg_h_score.size()).long(),self.use_cuda).float()
loss_name_KB=self.bceloss(pos_score,target)*3 +(self.bceloss(neg_h_score,target_neg)+self.bceloss(neg_t_score,target_neg)+self.bceloss(neg_r_score,target_neg)) * n_neg
# loss_name_KB=self.bceloss(pos_score,target)*3 +(self.bceloss(neg_h_score,target_neg)+self.bceloss(neg_t_score,target_neg)+self.bceloss(neg_r_score,target_neg))
loss+=loss_name_KB
# if self.mode=='joint':
# return loss_text+loss_KB+loss_Qtext+ loss_name_KB
# elif self.mode=='just_name_KB':
# return loss_text + loss_KB + loss_name_KB
# elif self.mode=='just_anchor':
# return loss_text + loss_KB + loss_Qtext
return loss
#