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
208
209
210
211
212
213
214
215
216
217
218
219
|
local THNN = require 'nn.THNN'
local SpatialFullConvolution, parent = torch.class('nn.SpatialFullConvolution','nn.Module')
function SpatialFullConvolution:__init(nInputPlane, nOutputPlane,
kW, kH, dW, dH, padW, padH, adjW, adjH)
parent.__init(self)
dW = dW or 1
dH = dH or 1
self.nInputPlane = nInputPlane
self.nOutputPlane = nOutputPlane
self.kW = kW
self.kH = kH
self.dW = dW
self.dH = dH
self.padW = padW or 0
self.padH = padH or 0
self.adjW = adjW or 0
self.adjH = adjH or 0
if self.adjW > self.dW - 1 or self.adjH > self.dH - 1 then
error('adjW and adjH must be smaller than self.dW - 1' ..
' and self.dH - 1 respectively')
end
self.weight = torch.Tensor(nInputPlane, nOutputPlane, kH, kW)
self.gradWeight = torch.Tensor(nInputPlane, nOutputPlane, kH, kW)
self.bias = torch.Tensor(self.nOutputPlane)
self.gradBias = torch.Tensor(self.nOutputPlane)
self.ones = torch.Tensor()
self:reset()
end
function SpatialFullConvolution:noBias()
self.bias = nil
self.gradBias = nil
return self
end
function SpatialFullConvolution:reset(stdv)
if stdv then
stdv = stdv * math.sqrt(3)
else
local nInputPlane = self.nInputPlane
local kH = self.kH
local kW = self.kW
stdv = 1/math.sqrt(kW*kH*nInputPlane)
end
self.weight:uniform(-stdv, stdv)
if self.bias then
self.bias:uniform(-stdv, stdv)
end
end
local function calculateAdj(targetSize, ker, pad, stride)
return (targetSize + 2 * pad - ker) % stride
end
function SpatialFullConvolution:backCompatibility()
self.adjW = self.adjW or 0
self.adjH = self.adjH or 0
end
function SpatialFullConvolution:updateOutput(input)
self:backCompatibility()
local inputTensor = input
local adjW, adjH = self.adjW, self.adjH
-- The input can be a table where the second element indicates the target
-- output size, in which case the adj factors are computed automatically
if type(inputTensor) == 'table' then
inputTensor = input[1]
local targetTensor = input[2]
local tDims = targetTensor:dim()
local tH = targetTensor:size(tDims-1)
local tW = targetTensor:size(tDims)
adjW = calculateAdj(tW, self.kW, self.padW, self.dW)
adjH = calculateAdj(tH, self.kH, self.padH, self.dH)
self.finput = self.finput or input[1].new()
self.fgradInput = self.fgradInput or input[1].new()
else
self.finput = self.finput or input.new()
self.fgradInput = self.fgradInput or input.new()
end
inputTensor.THNN.SpatialFullConvolution_updateOutput(
inputTensor:cdata(),
self.output:cdata(),
self.weight:cdata(),
THNN.optionalTensor(self.bias),
self.finput:cdata(),
self.fgradInput:cdata(),
self.kW, self.kH,
self.dW, self.dH,
self.padW, self.padH,
adjW, adjH
)
return self.output
end
function SpatialFullConvolution:updateGradInput(input, gradOutput)
self:backCompatibility()
if self.gradInput then
local inputTensor = input
local adjW, adjH = self.adjW, self.adjH
-- The input can be a table where the second element indicates the target
-- output size, in which case the adj factors are computed automatically
if type(inputTensor) == 'table' then
inputTensor = input[1]
local targetTensor = input[2]
local tDims = targetTensor:dim()
local tH = targetTensor:size(tDims-1)
local tW = targetTensor:size(tDims)
adjW = calculateAdj(tW, self.kW, self.padW, self.dW)
adjH = calculateAdj(tH, self.kH, self.padH, self.dH)
-- Momentarily extract the gradInput tensor
if type(self.gradInput) == 'table' then
self.gradInput = self.gradInput[1] or inputTensor.new()
end
end
inputTensor.THNN.SpatialFullConvolution_updateGradInput(
inputTensor:cdata(),
gradOutput:cdata(),
self.gradInput:cdata(),
self.weight:cdata(),
self.finput:cdata(),
self.kW, self.kH,
self.dW, self.dH,
self.padW, self.padH,
adjW, adjH
)
if type(input) == 'table' then
-- Create a zero tensor to be expanded and used as gradInput[2].
self.zeroScalar = self.zeroScalar or input[2].new(1):zero()
self.ones:resize(input[2]:dim()):fill(1)
local zeroTensor = self.zeroScalar
:view(table.unpack(self.ones:totable()))
:expandAs(input[2])
self.gradInput = {self.gradInput, zeroTensor}
end
return self.gradInput
end
end
function SpatialFullConvolution:accGradParameters(input, gradOutput, scale)
scale = scale or 1
self:backCompatibility()
local inputTensor = input
local adjW, adjH = self.adjW, self.adjH
-- The input can be a table where the second element indicates the target
-- output size, in which case the adj factors are computed automatically
if type(inputTensor) == 'table' then
inputTensor = input[1]
local targetTensor = input[2]
local tDims = targetTensor:dim()
local tH = targetTensor:size(tDims-1)
local tW = targetTensor:size(tDims)
adjW = calculateAdj(tW, self.kW, self.padW, self.dW)
adjH = calculateAdj(tH, self.kH, self.padH, self.dH)
end
inputTensor.THNN.SpatialFullConvolution_accGradParameters(
inputTensor:cdata(),
gradOutput:cdata(),
self.gradWeight:cdata(),
THNN.optionalTensor(self.gradBias),
self.finput:cdata(),
self.fgradInput:cdata(),
self.kW, self.kH,
self.dW, self.dH,
self.padW, self.padH,
adjW, adjH,
scale
)
end
function SpatialFullConvolution:type(type, tensorCache)
self.finput = self.finput and torch.Tensor()
self.fgradInput = self.fgradInput and torch.Tensor()
return parent.type(self, type, tensorCache)
end
function SpatialFullConvolution:__tostring__()
local s = string.format('%s(%d -> %d, %dx%d', torch.type(self),
self.nInputPlane, self.nOutputPlane, self.kW, self.kH)
if self.dW ~= 1 or self.dH ~= 1 or self.padW ~= 0 or self.padH ~= 0 then
s = s .. string.format(', %d,%d', self.dW, self.dH)
end
if (self.padW or self.padH) and (self.padW ~= 0 or self.padH ~= 0) then
s = s .. ', ' .. self.padW .. ',' .. self.padH
end
if (self.adjW or self.adjH) and (self.adjW ~= 0 or self.adjH ~= 0) then
s = s .. ', ' .. self.adjW .. ',' .. self.adjH
end
if self.bias then
return s .. ')'
else
return s .. ') without bias'
end
end
function SpatialFullConvolution:clearState()
nn.utils.clear(self, 'finput', 'fgradInput', '_input', '_gradOutput')
return parent.clearState(self)
end
|
