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
|
local VolumetricFullConvolution, parent
= torch.class('cudnn.VolumetricFullConvolution', 'nn.VolumetricFullConvolution')
local ffi = require 'ffi'
local find = require 'cudnn.find'
local errcheck = cudnn.errcheck
local checkedCall = find.checkedCall
local Convolution = cudnn.SpatialConvolution
-- if you change the configuration of the module manually, call this
function VolumetricFullConvolution:resetWeightDescriptors()
return Convolution.resetWeightDescriptors(
self,
{self.nInputPlane, self.nOutputPlane, self.kT, self.kH, self.kW}
)
end
function VolumetricFullConvolution:fastest(mode)
return Convolution.fastest(self, mode)
end
function VolumetricFullConvolution:setMode(fmode, bdmode, bwmode)
return Convolution.setMode(self, fmode, bdmode, bwmode)
end
function VolumetricFullConvolution:resetMode()
return Convolution.resetMode(self)
end
function VolumetricFullConvolution:createIODescriptors(input)
local batch = true
if input:dim() == 4 then
input = input:view(1, input:size(1), input:size(2),
input:size(3), input:size(4))
batch = false
end
assert(input:dim() == 5 and input:isContiguous());
self.iSize = self.iSize or torch.LongStorage(5):fill(0)
if Convolution.checkInputChanged(self, input) then
-- create input descriptor
local input_slice = input[{{},{1,self.nInputPlane},{},{}}]
self.iDesc = cudnn.toDescriptor(input_slice)
-- create conv descriptor
self.pad = {self.padT, self.padH, self.padW}
self.stride = {self.dT, self.dH, self.dW}
self.convDescData = { padA = self.pad, filterStrideA = self.stride,
dataType = cudnn.configmap(torch.type(self.weight))}
self.convDesc = cudnn.setConvolutionDescriptor(self.convDescData)
-- get output shape, resize output
local iwidth = input:size(5)
local iheight = input:size(4)
local idepth = input:size(3)
local owidth = (iwidth - 1) * self.dW - 2*self.padW + self.kW + self.adjW
local oheight = (iheight - 1) * self.dH - 2*self.padH + self.kH + self.adjH
local odepth = (idepth - 1) * self.dT - 2*self.padT + self.kT + self.adjT
local oSize = torch.IntTensor({input:size(1), self.nOutputPlane, odepth, oheight, owidth})
self.output:resize(oSize:long():storage())
-- create descriptor for output
local output_slice = self.output[{{},{1,self.nOutputPlane},{},{}}]
self.oDesc = cudnn.toDescriptor(output_slice)
self.oDescForBias = cudnn.toDescriptor(
self.output:view(self.output:size(1),
self.output:size(2),
self.output:size(3)*self.output:size(4),
self.output:size(5)))
self.input_offset = 0
self.output_offset = 0
self.weight_offset = 0
find:prepare(self, input_slice, output_slice)
if not batch then
self.output = self.output:view(self.output:size(2),
self.output:size(3),
self.output:size(4),
self.output:size(5))
end
end
end
local function makeContiguous(self, input, gradOutput)
if not input:isContiguous() then
self._input = self._input or input.new()
self._input:typeAs(input):resizeAs(input):copy(input)
input = self._input
end
if gradOutput and not gradOutput:isContiguous() then
self._gradOutput = self._gradOutput or gradOutput.new()
self._gradOutput:typeAs(gradOutput):resizeAs(gradOutput):copy(gradOutput)
gradOutput = self._gradOutput
end
return input, gradOutput
end
function VolumetricFullConvolution:updateOutput(input)
if not self.weightDesc then self:resetWeightDescriptors() end
self:createIODescriptors(input)
local finder = find.get()
-- Because SpatialFullConvolution is performing the adjoint of the forward
-- convolution operator, we need to swap the forward and backward passes.
local bwdDataAlgo = finder:backwardDataAlgorithm(self, {self.weightDesc[0], self.weight,
self.iDesc[0],self.input_slice,
self.convDesc[0], self.oDesc[0], self.output_slice})
local extraBuffer, extraBufferSize = cudnn.getSharedWorkspace()
checkedCall(self, 'cudnnConvolutionBackwardData', cudnn.getHandle(),
cudnn.scalar(input, 1),
self.weightDesc[0], self.weight:data(),
self.iDesc[0], input:data(),
self.convDesc[0], bwdDataAlgo,
extraBuffer, extraBufferSize,
cudnn.scalar(input, 0),
self.oDesc[0], self.output:data())
-- add bias
if self.bias then
errcheck('cudnnAddTensor', cudnn.getHandle(),
cudnn.scalar(input, 1), self.biasDesc[0], self.bias:data(),
cudnn.scalar(input, 1), self.oDescForBias[0], self.output:data())
end
return self.output
end
function VolumetricFullConvolution:updateGradInput(input, gradOutput)
if not self.gradInput then return end
self.gradInput:resizeAs(input)
assert(gradOutput:dim() == 4 or gradOutput:dim() == 5, 'gradOutput has to be 4D or 5D');
assert(gradOutput:isContiguous(), 'gradOutput has to be contiguous')
if not self.weightDesc then self:resetWeightDescriptors() end
self:createIODescriptors(input)
local finder = find.get()
local fwdAlgo = finder:forwardAlgorithm(self, {self.oDesc[0], self.output_slice,
self.weightDesc[0], self.weight,
self.convDesc[0], self.iDesc[0], self.input_slice})
local extraBuffer, extraBufferSize = cudnn.getSharedWorkspace()
checkedCall(self,'cudnnConvolutionForward', cudnn.getHandle(),
cudnn.scalar(input, 1),
self.oDesc[0], gradOutput:data(),
self.weightDesc[0], self.weight:data(),
self.convDesc[0],
fwdAlgo,
extraBuffer, extraBufferSize,
cudnn.scalar(input, 0),
self.iDesc[0], self.gradInput:data());
return self.gradInput
end
function VolumetricFullConvolution:accGradParameters(input, gradOutput, scale)
self.scaleT = self.scaleT or self.weight.new(1)
-- this line forces this member to always be on CPU (needed for cudnn)
self.scaleT = torch.type(self.weight) == 'torch.CudaDoubleTensor'
and self.scaleT:double() or self.scaleT:float()
scale = scale or 1.0
self.scaleT[1] = scale
input, gradOutput = makeContiguous(self, input, gradOutput)
assert(gradOutput:dim() == 4 or gradOutput:dim() == 5,
'gradOutput has to be a 4D or 5D tensor');
self:createIODescriptors(input)
if not self.weightDesc then self:resetWeightDescriptors() end
-- gradBias
local finder = find.get()
local bwdFilterAlgo = finder:backwardFilterAlgorithm(self, {self.oDesc[0], self.output_slice,
self.iDesc[0], self.input_slice,
self.convDesc[0], self.weightDesc[0], self.weight})
errcheck('cudnnConvolutionBackwardBias', cudnn.getHandle(),
self.scaleT:data(),
self.oDescForBias[0], gradOutput:data(),
cudnn.scalar(input, 1),
self.biasDesc[0], self.gradBias:data());
local extraBuffer, extraBufferSize = cudnn.getSharedWorkspace()
-- gradWeight
checkedCall(self, 'cudnnConvolutionBackwardFilter', cudnn.getHandle(),
self.scaleT:data(),
self.oDesc[0], gradOutput:data(),
self.iDesc[0], input:data(),
self.convDesc[0],
bwdFilterAlgo,
extraBuffer, extraBufferSize,
cudnn.scalar(input, 1),
self.weightDesc[0], self.gradWeight:data());
end
function VolumetricFullConvolution:clearDesc()
return Convolution.clearDesc(self)
end
function VolumetricFullConvolution:write(f)
self:clearDesc()
local var = {}
for k,v in pairs(self) do
var[k] = v
end
f:writeObject(var)
end
function VolumetricFullConvolution:clearState()
return Convolution.clearState(self)
end
|
