-
Notifications
You must be signed in to change notification settings - Fork 38
/
Copy pathgru_dec.py
157 lines (129 loc) · 6.19 KB
/
gru_dec.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
import torch
import torch as T
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import math
from utils_pg import *
class GRUAttentionDecoder(nn.Module):
def __init__(self, input_size, hidden_size, ctx_size, device, copy, coverage, is_predicting):
super(GRUAttentionDecoder, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.ctx_size = ctx_size
self.is_predicting = is_predicting
self.device = device
self.copy = copy
self.coverage = coverage
self.W = nn.Parameter(torch.Tensor(2 * self.hidden_size, self.input_size))
self.U = nn.Parameter(torch.Tensor(2 * self.hidden_size, self.hidden_size))
self.b = nn.Parameter(torch.Tensor(2 * self.hidden_size))
self.Wx = nn.Parameter(torch.Tensor(self.hidden_size, self.input_size))
self.Ux = nn.Parameter(torch.Tensor(self.hidden_size, self.hidden_size))
self.bx = nn.Parameter(torch.Tensor(self.hidden_size))
self.Wc_att = nn.Parameter(torch.Tensor(self.ctx_size, self.ctx_size))
self.b_att = nn.Parameter(torch.Tensor(self.ctx_size))
self.W_comb_att = nn.Parameter(torch.Tensor(self.ctx_size, self.hidden_size))
self.U_att = nn.Parameter(torch.Tensor(1, self.ctx_size))
self.U_nl = nn.Parameter(torch.Tensor(2 * self.hidden_size, self.hidden_size))
self.b_nl = nn.Parameter(torch.Tensor(2 * self.hidden_size))
self.Ux_nl = nn.Parameter(torch.Tensor(self.hidden_size, self.hidden_size))
self.bx_nl = nn.Parameter(torch.Tensor(self.hidden_size))
self.Wc = nn.Parameter(torch.Tensor(2 * self.hidden_size, self.ctx_size))
self.Wcx = nn.Parameter(torch.Tensor(self.hidden_size, self.ctx_size))
if self.coverage:
self.W_coverage= nn.Parameter(torch.Tensor(self.ctx_size, 1))
self.init_weights()
def init_weights(self):
init_ortho_weight(self.W)
init_ortho_weight(self.U)
init_bias(self.b)
init_ortho_weight(self.Wx)
init_ortho_weight(self.Ux)
init_bias(self.bx)
init_ortho_weight(self.Wc_att)
init_bias(self.b_att)
init_ortho_weight(self.W_comb_att)
init_ortho_weight(self.U_att)
init_ortho_weight(self.U_nl)
init_bias(self.b_nl)
init_ortho_weight(self.Ux_nl)
init_bias(self.bx_nl)
init_ortho_weight(self.Wc)
init_ortho_weight(self.Wcx)
if self.coverage:
init_ortho_weight(self.W_coverage)
def forward(self, y_emb, context, init_state, x_mask, y_mask, xid=None, init_coverage=None):
def _get_word_atten(pctx, h1, x_mask, acc_att=None): #acc_att: B * len(x)
if acc_att is not None:
h = F.linear(h1, self.W_comb_att) + F.linear(T.transpose(acc_att, 0, 1).unsqueeze(2), self.W_coverage) # len(x) * B * ?
else:
h = F.linear(h1, self.W_comb_att)
unreg_att = T.tanh(pctx + h) * x_mask
unreg_att = F.linear(unreg_att, self.U_att)
word_atten = T.exp(unreg_att - T.max(unreg_att, 0, keepdim = True)[0]) * x_mask
sum_word_atten = T.sum(word_atten, 0, keepdim = True)
word_atten = word_atten / sum_word_atten
return word_atten
def recurrence(x, xx, y_mask, pre_h, pctx, context, x_mask, acc_att=None):
tmp1 = T.sigmoid(F.linear(pre_h, self.U) + x)
r1, u1 = tmp1.chunk(2, 1)
h1 = T.tanh(F.linear(pre_h * r1, self.Ux) + xx)
h1 = u1 * pre_h + (1.0 - u1) * h1
h1 = y_mask * h1 + (1.0 - y_mask) * pre_h
# len(x) * batch_size * 1
if self.coverage:
word_atten = _get_word_atten(pctx, h1, x_mask, acc_att)
else:
word_atten = _get_word_atten(pctx, h1, x_mask)
atted_ctx = T.sum(word_atten * context, 0)
tmp2 = T.sigmoid(F.linear(atted_ctx, self.Wc) + F.linear(h1, self.U_nl) + self.b_nl)
r2, u2 = tmp2.chunk(2, 1)
h2 = T.tanh(F.linear(atted_ctx, self.Wcx) + F.linear(h1 * r2, self.Ux_nl) + self.bx_nl)
h2 = u2 * h1 + (1.0 - u2) * h2
h2 = y_mask * h2 + (1.0 - y_mask) * h1
word_atten_ = T.transpose(word_atten.view(x_mask.size(0), -1), 0, 1) # B * len(x)
if self.coverage:
acc_att += word_atten_
return h2, h2, atted_ctx, word_atten_, acc_att
else:
return h2, h2, atted_ctx, word_atten_
hs, ss, atts, dists, xids, cs = [], [], [], [], [], []
hidden = init_state
acc_att = init_coverage
if self.copy:
xid = T.transpose(xid, 0, 1) # B * len(x)
pctx = F.linear(context, self.Wc_att, self.b_att)
x = F.linear(y_emb, self.W, self.b)
xx = F.linear(y_emb, self.Wx, self.bx)
steps = range(y_emb.size(0))
for i in steps:
if self.coverage:
cs += [acc_att]
hidden, s, att, att_dist, acc_att = recurrence(x[i], xx[i], y_mask[i], hidden, pctx, context, x_mask, acc_att)
else:
hidden, s, att, att_dist = recurrence(x[i], xx[i], y_mask[i], hidden, pctx, context, x_mask)
hs += [hidden]
ss += [s]
atts += [att]
dists += [att_dist]
xids += [xid]
if self.coverage:
if self.is_predicting :
cs += [acc_att]
cs = cs[1:]
cs = T.stack(cs).view(y_emb.size(0), *cs[0].size())
hs = T.stack(hs).view(y_emb.size(0), *hs[0].size())
ss = T.stack(ss).view(y_emb.size(0), *ss[0].size())
atts = T.stack(atts).view(y_emb.size(0), *atts[0].size())
dists = T.stack(dists).view(y_emb.size(0), *dists[0].size())
if self.copy:
xids = T.stack(xids).view(y_emb.size(0), *xids[0].size())
if self.copy and self.coverage:
return hs, ss, atts, dists, xids, cs
elif self.copy:
return hs, ss, atts, dists, xids
elif self.coverage:
return hs, ss, atts, dists, cs
else:
return hs, ss, atts