Integrated CUDA conv routines from cuda-convnet library.

3 files changed, 81 insertions, 0 deletions

diff --git a/SpatialConvolutionCUDA.lua b/SpatialConvolutionCUDA.lua
new file mode 100644
index 0000000..ccc0284
--- /dev/null
+++ b/SpatialConvolutionCUDA.lua
@@ -0,0 +1,50 @@
+local SpatialConvolutionCUDA, parent = torch.class('nn.SpatialConvolutionCUDA', 'nn.Module')
+
+function SpatialConvolutionCUDA:__init(nInputPlane, nOutputPlane, kW, kH, dW, dH)
+   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.weight = torch.Tensor(nInputPlane, kH, kW, nOutputPlane)
+   self.gradWeight = torch.Tensor(nInputPlane, kH, kW, nOutputPlane)
+   
+   self:reset()
+end
+
+function SpatialConvolutionCUDA:reset(stdv)
+   if stdv then
+      stdv = stdv * math.sqrt(3)
+   else
+      stdv = 1/math.sqrt(self.kW*self.kH*self.nInputPlane)
+   end
+   self.weight:uniform(-stdv, stdv)
+end
+
+function SpatialConvolutionCUDA:updateOutput(input)
+   return input.nn.SpatialConvolutionCUDA_updateOutput(self, input)
+end
+
+function SpatialConvolutionCUDA:updateGradInput(input, gradOutput)
+   return input.nn.SpatialConvolutionCUDA_updateGradInput(self, input, gradOutput)
+end
+
+function SpatialConvolutionCUDA:accGradParameters(input, gradOutput, scale)
+   return input.nn.SpatialConvolutionCUDA_accGradParameters(self, input, gradOutput, scale)
+end
+
+function SpatialConvolutionCUDA:copy(sc)
+   local weight = sc.weight:clone()
+   weight:resize(sc.nOutputPlane, sc.nInputPlane * sc.kH * sc.kW)
+   weight = weight:t():contiguous()
+   weight:resize(sc.nInputPlane, sc.kH, sc.kW, sc.nOutputPlane)
+   self.weight:copy(weight)
+end
+
diff --git a/Transpose.lua b/Transpose.lua
new file mode 100644
index 0000000..a43729b
--- /dev/null
+++ b/Transpose.lua
@@ -0,0 +1,29 @@
+local Transpose, parent = torch.class('nn.Transpose', 'nn.Module')
+
+-- transpose dimensions:
+-- n = nn.Transpose({1,4},{1,3})
+-- will transpose dims 1 and 4, then 1 and 3...
+
+function Transpose:__init(...)
+   parent.__init(self)
+   self.permutations = {...}
+end
+
+function Transpose:updateOutput(input)
+   for _,perm in ipairs(self.permutations) do
+      input = input:transpose(perm[1],perm[2])
+   end
+   self.output:resizeAs(input):copy(input)
+   return self.output
+end
+
+function Transpose:updateGradInput(input, gradOutput)
+   local ndim = gradOutput:nDimension()
+   for i = #self.permutations,1,-1 do
+      local perm = self.permutations[i]
+      gradOutput = gradOutput:transpose(perm[1],perm[2])
+   end
+   self.gradInput:resizeAs(gradOutput):copy(gradOutput)
+   return self.gradInput
+end
+
diff --git a/init.lua b/init.lua
index 587412f..c6e9df9 100644
--- a/init.lua
+++ b/init.lua
@@ -13,6 +13,7 @@ include('Reshape.lua')
 include('Select.lua')
 include('Narrow.lua')
 include('Replicate.lua')
+include('Transpose.lua')
 
 include('Copy.lua')
 include('Min.lua')
@@ -59,6 +60,7 @@ include('SpatialConvolution.lua')
 include('SpatialFullConvolution.lua')
 include('SpatialFullConvolutionMap.lua')
 include('SpatialConvolutionMM.lua')
+include('SpatialConvolutionCUDA.lua')
 include('SpatialConvolutionMap.lua')
 include('SpatialSubSampling.lua')
 include('SpatialMaxPooling.lua')