neural_decompiler/model_attention.py at master · satos---jp/neural_decompiler · GitHub

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
import numpy as np
import chainer
import chainer.links as L
import chainer.functions as F

# from https://arxiv.org/pdf/1409.0473.pdf. Global attention.
class Attention(chainer.Chain):
	def __init__(self, n_units):
		super(Attention, self).__init__()
		with self.init_scope():
			self.att_W = L.Linear(n_units, n_units)
			self.att_U = L.Linear(n_units*2, n_units)
			self.att_v = L.Linear(n_units, 1)

		self.n_units = n_units
		self.Uhxs = None

	#@profile
	def init_hxs(self,hxs):
		self.hxs = hxs
		self.Uhxs = self.att_U(hxs)
		self.Uhxs_len = hxs.shape[0]

	#@profile
	def forward(self,nhx):
		nhx_for_att = nhx[-1]
		Wnhx = self.att_W(nhx_for_att)
		#print(Wnhx.shape,Uhxs.shape)
		#print((Wnhx + Uhxs).shape)

		att_rates = F.softmax(F.reshape(self.att_v(
				F.tanh(
					Wnhx + self.Uhxs
				)
		),(1,self.Uhxs_len)))
		#print(hxs.shape,att_rates.shape)
		#print(att_rates)
		#print(F.matmul(att_rates,hxs).shape)
		ctx = F.matmul(att_rates,self.hxs)[0]
		#print(ctx.shape)
		return ctx

	def reset(self):
		self.hxs = None
		self.Uhxs = None


class DotAttention(chainer.Chain):
	def __init__(self, n_units):
		super(DotAttention, self).__init__()
		self.n_units = n_units
		self.hxs = None

	#@profile
	def forward(self,xh,yh):
		#print(xh.shape,yh.shape)

		att_rates = F.softmax(F.matmul(yh,xh.T))
		#print(hxs.shape,att_rates.shape)
		#print(att_rates.shape)
		ds = F.sum(att_rates,axis=1)
		#print(ds[:10],ds.shape)
		#print(F.matmul(att_rates,hxs).shape)
		ctx = F.tanh(F.matmul(att_rates,xh))
		#print(ctx.shape)
		return ctx

class GlobalGeneralAttention(chainer.Chain):
	def __init__(self, n_units):
		super(GlobalGeneralAttention, self).__init__()
		self.n_units = n_units
		self.hxs = None
		with self.init_scope():
			self.Wa = L.Linear(n_units*2, n_units*2)

	#@profile
	def forward(self,xh,yh):
		#print(xh.shape,yh.shape)

		# O([a,b] * [b,c]) = a*b*c
		# a*b*c+a*c*d = a*c*(b+d)
		# a*b*d+b*c*d = (a+c)*b*d

		att_rates = F.softmax(F.matmul(self.Wa(yh),xh.T))
		#print(hxs.shape,att_rates.shape)
		#print(att_rates.shape)
		ds = F.sum(att_rates,axis=1)
		#print(ds[:10],ds.shape)
		#print(F.matmul(att_rates,hxs).shape)
		ctx = F.tanh(F.matmul(att_rates,xh))
		#print(ctx.shape)
		return ctx

class LocalDotAttention(chainer.Chain):
	def __init__(self, n_units):
		super(DotAttention, self).__init__()
		self.n_units = n_units
		self.hxs = None
		with self.init_scope():
			self.Wp = L.Linear(n_units, n_units)
			self.vp = L.Linear(n_units, 1)

	#@profile
	def forward(self,xh,yh):
		#print(xh.shape,yh.shape)

		p_t = xh.shape[0] * F.sigmoid(self.vp(F.tanh(self.Wp(yh))))
		#p_att = F.exp(
		att_rates = F.softmax(F.matmul(yh,xh.T)*p_att)

		#print(hxs.shape,att_rates.shape)
		#print(att_rates.shape)
		ds = F.sum(att_rates,axis=1)
		#print(ds[:10],ds.shape)
		#print(F.matmul(att_rates,hxs).shape)
		ctx = F.tanh(F.matmul(att_rates,xh))
		#print(ctx.shape)
		return ctx