Merge pull request #171 from szagoruyko/R5-gradInput-move

[R5] Move gradInput:resizeAs(input) everywhere

9 files changed, 19 insertions, 37 deletions

diff --git a/BatchNormalization.lua b/BatchNormalization.lua
index a342063..83597d3 100644
--- a/BatchNormalization.lua
+++ b/BatchNormalization.lua
@@ -48,7 +48,6 @@ function BatchNormalization:createIODescriptors(input)
       local nFeature = self.running_mean:numel()
       self.iSize = input:size()
       self.output:resizeAs(input)
-      self.gradInput:resizeAs(input)
       self.iDesc = cudnn.toDescriptor(input)
       self.oDesc = cudnn.toDescriptor(self.output)
       local biasSize = torch.ones(self.nDim):totable()
@@ -88,6 +87,7 @@ end
 local function backward(self,input,gradOutput, scale)
    assert(gradOutput:isContiguous())
    self:createIODescriptors(input)
+   self.gradInput:resizeAs(input)
    scale = scale or 1
    scaleTens:fill(scale)
    errcheck('cudnnBatchNormalizationBackward',
diff --git a/Pointwise.lua b/Pointwise.lua
index 92b3e45..93298ad 100644
--- a/Pointwise.lua
+++ b/Pointwise.lua
@@ -12,7 +12,6 @@ function Pointwise:createIODescriptors(input)
    assert(self.mode, 'mode is not set. (trying to use base class?)');
    assert(input:isContiguous(), 'Non-contiguous inputs not supported yet');
    if not self.inplace then
-       self.gradInput:resizeAs(input)
        self.output:resizeAs(input)
    end
 
@@ -60,7 +59,12 @@ function Pointwise:updateGradInput(input, gradOutput)
       gradOutput = self._gradOutput
    end
    self:createIODescriptors(input)
-   if self.inplace then self.output:set(input); self.gradInput:set(gradOutput) end
+   if self.inplace then
+      self.output:set(input);
+      self.gradInput:set(gradOutput)
+   else
+      self.gradInput:resizeAs(input)
+   end
    errcheck('cudnnActivationBackward',
             cudnn.getHandle(), self.activDesc[0],
             one:data(),
diff --git a/Pooling.lua b/Pooling.lua
index d004563..45afccb 100644
--- a/Pooling.lua
+++ b/Pooling.lua
@@ -53,9 +53,6 @@ function Pooling:createIODescriptors(input)
       input:size(1) ~= self.iSize[1] or input:size(2) ~= self.iSize[2]
    or input:size(3) ~= self.iSize[3] or input:size(4) ~= self.iSize[4] then
       self.iSize = input:size()
-      -- resize gradInput
-      self.gradInput:resizeAs(input)
-      -- resize output
       local oW, oH
       if self.ceil_mode then
          oW = math.ceil((input:size(4)+self.padW*2 - self.kW)/self.dW + 1)
@@ -70,9 +67,6 @@ function Pooling:createIODescriptors(input)
       self.iDesc = cudnn.toDescriptor(input)
       self.oDesc = cudnn.toDescriptor(self.output)
       if not batch then
-         self.gradInput = self.gradInput:view(self.gradInput:size(2),
-                                              self.gradInput:size(3),
-                                              self.gradInput:size(4))
          self.output = self.output:view(self.output:size(2),
                                         self.output:size(3),
                                         self.output:size(4))
@@ -102,6 +96,7 @@ function Pooling:updateGradInput(input, gradOutput)
       self._gradOutput:resizeAs(gradOutput):copy(gradOutput)
       gradOutput = self._gradOutput
    end
+   self.gradInput:resizeAs(input)
    if not self.poolDesc then self:resetPoolDescriptors() end
    self:createIODescriptors(input)
    errcheck('cudnnPoolingBackward',
diff --git a/Pooling3D.lua b/Pooling3D.lua
index cce67c3..e4c0218 100644
--- a/Pooling3D.lua
+++ b/Pooling3D.lua
@@ -58,8 +58,6 @@ function Pooling:createIODescriptors(input)
    or input:size(3) ~= self.iSize[3] or input:size(4) ~= self.iSize[4]
    or input:size(5) ~= self.iSize[5] then
       self.iSize = input:size()
-      -- resize gradInput
-      self.gradInput:resizeAs(input)
       -- resize output
       local oW, oH, oT
       if self.ceil_mode then
@@ -77,10 +75,6 @@ function Pooling:createIODescriptors(input)
       self.iDesc = cudnn.toDescriptor(input)
       self.oDesc = cudnn.toDescriptor(self.output)
       if not batch then
-         self.gradInput = self.gradInput:view(self.gradInput:size(2),
-                                              self.gradInput:size(3),
-                                              self.gradInput:size(4),
-                                              self.gradInput:size(5))
          self.output = self.output:view(self.output:size(2),
                                         self.output:size(3),
                                         self.output:size(4),
@@ -105,6 +99,9 @@ function Pooling:updateOutput(input)
 end
 
 function Pooling:updateGradInput(input, gradOutput)
+   if not self.gradInput then return end
+   self.gradInput:resizeAs(input)
+
    assert(gradOutput:dim() == 4 or gradOutput:dim() == 5);
    if not gradOutput:isContiguous() then
       self._gradOutput = self._gradOutput or gradOutput.new()
diff --git a/SpatialConvolution.lua b/SpatialConvolution.lua
index b92dd57..1fd3ea0 100644
--- a/SpatialConvolution.lua
+++ b/SpatialConvolution.lua
@@ -388,8 +388,8 @@ end
 
 function SpatialConvolution:updateGradInput(input, gradOutput)
     if not self.gradInput then return end
-
     self.gradInput:resizeAs(input)
+
     input, gradOutput = makeContiguous(self, input, gradOutput)
     assert(gradOutput:dim() == 3 or gradOutput:dim() == 4, 'gradOutput has to be 3D or 4D');
     if not self.weightDesc then self:resetWeightDescriptors() end
diff --git a/SpatialCrossMapLRN.lua b/SpatialCrossMapLRN.lua
index f6e7cd9..1f4ba33 100644
--- a/SpatialCrossMapLRN.lua
+++ b/SpatialCrossMapLRN.lua
@@ -36,15 +36,11 @@ function LRN:createIODescriptors(input)
       input:size(1) ~= self.iSize[1] or input:size(2) ~= self.iSize[2]
    or input:size(3) ~= self.iSize[3] or input:size(4) ~= self.iSize[4] then
       self.iSize = input:size()
-      self.gradInput:resizeAs(input)
       self.output:resizeAs(input)
 
       -- create input/output descriptor
       self.iDesc = cudnn.toDescriptor(input)
       if not batch then
-         self.gradInput = self.gradInput:view(self.gradInput:size(2),
-                                              self.gradInput:size(3),
-                                              self.gradInput:size(4))
          self.output = self.output:view(self.output:size(2),
                                         self.output:size(3),
                                         self.output:size(4))
@@ -70,6 +66,9 @@ function LRN:updateOutput(input)
 end
 
 function LRN:updateGradInput(input, gradOutput)
+   if not self.gradInput then return end
+   self.gradInput:resizeAs(input)
+
    assert(gradOutput:dim() == 3 or gradOutput:dim() == 4);
    if not gradOutput:isContiguous() then
       self._gradOutput = self._gradOutput or gradOutput.new()
diff --git a/SpatialFullConvolution.lua b/SpatialFullConvolution.lua
index 887e85b..cfbd61d 100644
--- a/SpatialFullConvolution.lua
+++ b/SpatialFullConvolution.lua
@@ -75,8 +75,6 @@ function SpatialFullConvolution:createIODescriptors(input)
     or input:size(3) ~= self.iSize[3] or input:size(4) ~= self.iSize[4] then
         self.iSize = input:size()
 
-        -- resize gradInput
-        if self.gradInput then self.gradInput:resizeAs(input); end
         assert(self.nInputPlane == input:size(2), 'input has to contain: '
                    .. self.nInputPlane
                    .. ' feature maps, but received input of size: '
@@ -291,9 +289,6 @@ function SpatialFullConvolution:createIODescriptors(input)
         end
 
         if not batch then
-            self.gradInput = self.gradInput:view(self.gradInput:size(2),
-                                                 self.gradInput:size(3),
-                                                 self.gradInput:size(4))
             self.output = self.output:view(self.output:size(2),
                                            self.output:size(3),
                                            self.output:size(4))
@@ -329,6 +324,7 @@ end
 
 function SpatialFullConvolution:updateGradInput(input, gradOutput)
     if not self.gradInput then return end
+    self.gradInput:resizeAs(input)
 
     assert(gradOutput:dim() == 3 or gradOutput:dim() == 4, 'gradOutput has to be 3D or 4D');
     assert(gradOutput:isContiguous(), 'gradOutput has to be contiguous')
diff --git a/SpatialSoftMax.lua b/SpatialSoftMax.lua
index 4b3a488..167eb1f 100644
--- a/SpatialSoftMax.lua
+++ b/SpatialSoftMax.lua
@@ -35,22 +35,16 @@ function SpatialSoftMax:createIODescriptors(input)
       input:size(1) ~= self.iSize[1] or input:size(2) ~= self.iSize[2]
    or input:size(3) ~= self.iSize[3] or input:size(4) ~= self.iSize[4] then
       self.iSize = input:size()
-      self.gradInput:resizeAs(input)
       self.output:resizeAs(input)
       self.iDesc = cudnn.toDescriptor(input)
       self.oDesc = cudnn.toDescriptor(self.output)
       if not singleDim and not batch then
-         self.gradInput = self.gradInput:view(self.gradInput:size(2),
-                                              self.gradInput:size(3),
-                                              self.gradInput:size(4))
          self.output = self.output:view(self.output:size(2),
                                         self.output:size(3),
                                         self.output:size(4))
       elseif singleDim and not batch then
-         self.gradInput = self.gradInput:view(self.gradInput:size(2))
          self.output = self.output:view(self.output:size(2))
       elseif singleDim and batch then
-         self.gradInput = self.gradInput:view(self.gradInput:size(1), self.gradInput:size(2))
          self.output = self.output:view(self.output:size(1), self.output:size(2))
       end
    end
@@ -72,6 +66,7 @@ function SpatialSoftMax:updateOutput(input)
 end
 
 function SpatialSoftMax:updateGradInput(input, gradOutput)
+   self.gradInput:resizeAs(input)
    if not gradOutput:isContiguous() then
       self._gradOutput = self._gradOutput or gradOutput.new()
       self._gradOutput:resizeAs(gradOutput):copy(gradOutput)
diff --git a/VolumetricConvolution.lua b/VolumetricConvolution.lua
index e9efb64..fd5e9c7 100644
--- a/VolumetricConvolution.lua
+++ b/VolumetricConvolution.lua
@@ -76,8 +76,6 @@ function VolumetricConvolution:createIODescriptors(input)
    or input:size(3) ~= self.iSize[3] or input:size(4) ~= self.iSize[4]
    or input:size(5) ~= self.iSize[5] then
          self.iSize = input:size()
-         -- resize gradInput
-         if self.gradInput then self.gradInput:resizeAs(input); end
          -- create input descriptor
          self.iDesc = cudnn.toDescriptor(input)
          -- create conv descriptor
@@ -287,10 +285,6 @@ function VolumetricConvolution:createIODescriptors(input)
         -----------------------------------------------------------------------
 
          if not batch then
-            self.gradInput = self.gradInput:view(self.gradInput:size(2),
-                                                 self.gradInput:size(3),
-                                                 self.gradInput:size(4),
-                                                 self.gradInput:size(5))
             self.output = self.output:view(self.output:size(2),
                                            self.output:size(3),
                                            self.output:size(4),
@@ -337,6 +331,8 @@ end
 
 function VolumetricConvolution:updateGradInput(input, gradOutput)
    if not self.gradInput then return end
+   self.gradInput:resizeAs(input)
+
    input, gradOutput = makeContiguous(self, input, gradOutput)
    assert(gradOutput:dim() == 4 or gradOutput:dim() == 5,
           'gradOutput has to be a 4D or 5D tensor');