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
|
local CAdd, parent = torch.class("nn.CAdd", "nn.Module")
function CAdd:__init(...)
parent.__init(self)
local arg = {...}
self.size = torch.LongStorage()
local n = #arg
if n == 1 and torch.type(arg[1]) == 'torch.LongStorage' then
self.size:resize(#arg[1]):copy(arg[1])
else
self.size:resize(n)
for i=1,n do
self.size[i] = arg[i]
end
end
self.bias = torch.Tensor(self.size)
self.gradBias = torch.Tensor(self.size)
self.output:resize(self.size)
self:reset()
end
function CAdd:reset(stdv)
if stdv then
--std of uniform distribution on interval [-a,a] = a/sqrt(3)
stdv = stdv * math.sqrt(3)
else
stdv = 1.0/math.sqrt(self.bias:nElement())
end
self.bias:uniform(-stdv,stdv)
end
function CAdd:updateOutput(input)
self._output = self._output or input.new()
self._bias = self._bias or input.new()
self._expand = self._expand or input.new()
self._repeat = self._repeat or input.new()
self.output:resizeAs(input):copy(input)
if input:nElement() == self.bias:nElement() then
self.output:add(self.bias)
else
if self.bias:dim() == input:dim() then
self._output:set(self.output)
self._bias:set(self.bias)
else
local batchSize = input:size(1)
self._output:view(self.output, batchSize, -1)
self._bias:view(self.bias, 1, -1)
end
self._expand:expandAs(self._bias, self._output)
--expandAs uses stride 0 and self._expand is not contiguous
--cuda ops may assume contiguous input
if torch.type(input) == 'torch.CudaTensor' then
self._repeat:resizeAs(self._expand):copy(self._expand)
self._output:add(self._repeat)
else
self._output:add(self._expand)
end
end
return self.output
end
function CAdd:updateGradInput(input, gradOutput)
self.gradInput = self.gradInput or input.new()
self.gradInput:resizeAs(gradOutput):copy(gradOutput)
return self.gradInput
end
function CAdd:accGradParameters(input, gradOutput, scale)
scale = scale or 1
self._gradBias = self._gradBias or gradOutput.new()
self._gradOutput = self._gradOutput or gradOutput.new()
self._repeat = self._repeat or gradOutput.new()
if self.bias:nElement() == gradOutput:nElement() then
self.gradBias:add(scale, gradOutput)
else
if self.bias:dim() == gradOutput:dim() then
self._gradBias:set(self.gradBias)
self._gradOutput:set(gradOutput)
else
local batchSize = input:size(1)
self._gradBias:view(self.gradBias, 1, -1)
self._gradOutput:view(gradOutput, batchSize, -1)
end
self._gradBias:expandAs(self._gradBias, self._gradOutput)
--expandAs uses stride 0 and self._gradBias is not contiguous
--cuda ops may assume contiguous input
if torch.type(self._gradBias) == 'torch.CudaTensor' then
self._repeat:resizeAs(self._gradBias):copy(self._gradBias)
self._repeat:add(scale, self._gradOutput)
self._gradBias:copy(self._repeat)
else
self._gradBias:add(scale, self._gradOutput)
end
end
end
function CAdd:type(type, tensorCache)
if type then
self:clearState()
end
return parent.type(self, type, tensorCache)
end
function CAdd:clearState()
nn.utils.clear(self, {
'_gradBias',
'_expand',
'_output',
'_bias',
'_repeat'
})
return parent.clearState(self)
end
|
