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
|
local Cosine, parent = torch.class('nn.Cosine', 'nn.Module')
function Cosine:__init(inputSize,outputSize)
parent.__init(self)
self.weight = torch.Tensor(outputSize,inputSize)
self.gradWeight = torch.Tensor(outputSize,inputSize)
self:reset()
end
function Cosine:reset(stdv)
if stdv then
stdv = stdv * math.sqrt(3)
else
stdv = 1./math.sqrt(self.weight:size(1))
end
self.weight:uniform(-stdv, stdv)
end
function Cosine:updateOutput(input)
local inputSize = self.weight:size(2)
local outputSize = self.weight:size(1)
self._weightNorm = self._weightNorm or self.weight.new()
self._inputNorm = self._inputNorm or self.weight.new()
-- y_j = (w_j * x) / ( || w_j || * || x || )
self._weightNorm:norm(self.weight,2,2):add(1e-12)
if input:dim() == 1 then
self.output:resize(outputSize):zero()
self.output:addmv(1, self.weight, input)
self.__norm = input:norm()+1e-12
self.output:cdiv(self._weightNorm:view(outputSize)):div(self.__norm)
elseif input:dim() == 2 then
local batchSize = input:size(1)
local nElement = self.output:nElement()
self.output:resize(batchSize, outputSize)
if self.output:nElement() ~= nElement then
self.output:zero()
end
self.output:addmm(0, self.output, 1, input, self.weight:t())
self._inputNorm:norm(input,2,2):add(1e-12)
self.output:cdiv(self._weightNorm:view(1,outputSize):expandAs(self.output))
self.output:cdiv(self._inputNorm:expandAs(self.output))
else
error('input must be vector or matrix')
end
return self.output
end
function Cosine:updateGradInput(input, gradOutput)
if not self.gradInput then
return
end
local inputSize = self.weight:size(2)
local outputSize = self.weight:size(1)
--[[
dy_j w_ji x_i
---- = ------------------- - y_j ---------
dx_i || w_j || * || x || || x ||^2
--]]
local nElement = self.gradInput:nElement()
self.gradInput:resizeAs(input)
if self.gradInput:nElement() ~= nElement then
self.gradInput:zero()
end
if input:dim() == 1 then
self._weight = self._weight or input.new()
self._weight:resizeAs(self.weight):copy(self.weight)
self._weight:cdiv(self._weightNorm:expandAs(self.weight))
self._weight:div(self.__norm)
self._weight:addr(1, self._weight, -1/(self.__norm*self.__norm), self.output, input)
self.gradInput:addmv(0, 1, self._weight:t(), gradOutput)
elseif input:dim() == 2 then
local inputNorm = self._inputNorm:expandAs(input)
local weightNorm = self._weightNorm:view(1,outputSize):expandAs(gradOutput)
self.gradInput:copy(input):cdiv(inputNorm)
self._gradOutput = self._gradOutput or gradOutput.new()
self._gradOutput:resizeAs(gradOutput):copy(gradOutput)
self._gradOutput:cmul(self.output)
self._sum = self._sum or input.new()
self._sum:sum(self._gradOutput, 2)
self.gradInput:cmul(self._sum:expandAs(input))
self._gradOutput:resizeAs(gradOutput):copy(gradOutput)
self._gradOutput:cdiv(weightNorm)
self.gradInput:addmm(-1, self.gradInput, 1, self._gradOutput, self.weight)
self.gradInput:cdiv(inputNorm)
end
return self.gradInput
end
function Cosine:accGradParameters(input, gradOutput, scale)
scale = scale or 1
local inputSize = self.weight:size(2)
local outputSize = self.weight:size(1)
--[[
dy_j x_i w_ji
----- = ------------------- - y_j -----------
dw_ji || w_j || * || x || || w_j ||^2
--]]
if input:dim() == 1 then
self._gradOutput = self._gradOutput or gradOutput.new()
self._gradOutput:resizeAs(gradOutput):copy(gradOutput)
local weightNorm = self._weightNorm:view(outputSize)
self._gradOutput:cdiv(weightNorm)
self.gradWeight:addr(scale/self.__norm, self._gradOutput, input)
self._gradOutput:cdiv(weightNorm)
self._gradOutput:cmul(self.output)
self._weight = self._weight or self.weight.new()
self._weight:resizeAs(self._weight):copy(self.weight)
self._weight:cmul(self._gradOutput:view(outputSize, 1):expandAs(self.weight))
self.gradWeight:add(-1, self._weight)
elseif input:dim() == 2 then
self._weight = self._weight or self.weight.new()
self._weight:resizeAs(self.weight):copy(self.weight)
self._gradOutput = self._gradOutput or gradOutput.new()
self._gradOutput:resizeAs(gradOutput):copy(gradOutput)
self._gradOutput:cmul(self.output)
self._sum = self._sum or input.new()
self._sum:sum(self._gradOutput, 1)
local grad = self._sum[1]
grad:cdiv(self._weightNorm:select(2,1))
self._weight:cmul(grad:view(outputSize,1):expandAs(self._weight))
local input_ = self._gradOutput
input_:resizeAs(input):copy(input)
input_:cdiv(self._inputNorm:expandAs(input))
self._weight:addmm(-1, self._weight, 1, gradOutput:t(), input_)
self._weight:cdiv(self._weightNorm:expandAs(self._weight))
self.gradWeight:add(self._weight)
else
error"1D or 2D input expected"
end
end
function Cosine:type(type, tensorCache)
if type then
-- prevent premature memory allocations
self._input = nil
self._weight = nil
self._inputNorm = nil
self._weightNorm = nil
self._gradOutput = nil
self._sum = nil
end
return parent.type(self, type, tensorCache)
end
function Cosine:clearState()
nn.utils.clear(self, {
'_input',
'_weight',
'_gradOutput',
'_sum',
'_inputNorm',
'_weightNorm',
})
return parent.clearState(self)
end
|
