Any success in applying this to VGG model?

[sheikirfanbasha]

I have been trying to apply the logic of deconvolution to VGG model. But so far have no success. The following is the modified code:

   local feature_map = state.network.modules[state.network_table[currentId].i].output[1][filterNum]:clone()
   print("size of feature map is:")
   print(feature_map:size())
   state.network.modules[state.network_table[currentId].i].output[1]:fill(0)
   state.network.modules[state.network_table[currentId].i].output[1][filterNum]:copy(feature_map)

   --- Guided back-propogation of ReLU
   -- http://arxiv.org/pdf/1412.6806v3.pdf
   local currentGradOutput = state.network.modules[state.network_table[currentId].i].output
   local currentModule = state.network.modules[state.network_table[currentId].i]
   for i=state.network_table[currentId].i-1,1,-1 do
     print("previous model: "..state.network.modules[i].__typename) 
     local previousModule = state.network.modules[i]
     if currentModule.__typename =="nn.ReLU" then 
        currentGradOutput = currentModule:backward(previousModule.output, currentGradOutput)
        currentGradOutput = currentModule:forward(currentGradOutput)
     else
        currentGradOutput = currentModule:backward(previousModule.output, currentGradOutput)
     end
     currentModule.gradInput = currentGradOutput
     currentModule = previousModule
   end
   currentGradOutput = currentModule:backward(input_img, currentGradOutput)
   print(currentGradOutput:size())
   local gra = image.toDisplayTensor{input=currentGradOutput[1],
               zoom = 8}
   -- local fea = image.toDisplayTensor{input=feature_map,
   --             zoom = 8}
   image.save('vgg_deconv.png', gra)

Error

bad argument #5 to 'v' (2D or 4D weight tensor expected, but got: [] at /torch/extra/nn/lib/THNN/generic/SpatialConvolutionMM.c:15)