SpatialDepthWiseConvolution swaps height and width

[diovisgood]

I've met a problem with SpatialDepthWiseConvolution
When I pass image, say 4x5 (height=4, width=5), this layer outputs image 5x4.
Proof code:

-- Make model:
local model = nn.SpatialDepthWiseConvolution(3, 1, 3, 3, 1, 1, 1, 1)
-- Generate input: batch = 2, nChannels = 3, height = 4, width = 5
local X = torch.randn(2, 3, 4, 5)
print('X:\n'..tostring(X))
-- Pass input through SpatialDepthWiseConvolution
local Yhat = model:forward(X)
print('Yhat:\n'..tostring(Yhat))

This script outputs:

X: ...
[torch.DoubleTensor of size 2x3x4x5]

Yhat: ...
[torch.DoubleTensor of size 2x3x5x4]

This breaks all the sequence of NN, and model does not work as intended!
Also, this contradicts to documentation, that says:

The most noticiable difference is the output dimension of SpatialConvolution is nOutputPlane x oheight x owidth, while for SpatialDepthWiseConvolution it is (nOutputPlane x nInputPlane) x oheight x owidth.

[diovisgood]

Unfortunately, I have no time to dig into C-code of THNN implementation of SpatialDepthWiseConvolution.
There is a way to emulate this module using containers Concat and Parallel along with standard SpatialConvolution module:

  local depth_wise_conv = nn.Concat(2)
  for o = 1, nOutputPlane do
    local out = nn.Parallel(2, 2)
    for i = 1, nInputPlane do
      local seq = nn.Sequential()
      local conv = nn.SpatialConvolution(1, 1, kW, kH, dW, dH, pW, pH):noBias()
      conv.weight:copy(weight[o][i])
      seq:add( nn.Reshape(1, inputHeight, inputWidth) )
      seq:add( conv )
      out:add( seq )
    end
    depth_wise_conv:add( out )
  end

But I've spend some time and created a module in pure lua that performs spatial depth wise convolution and (I hope) works as intended. You may find source code and unit tests here:
SpatialDepthWiseConvolution in torch (lua)

I'll be waiting for someone to fix this problem, because low-level implementation should be faster!

[gkanno]

How about these ?
gkanno@05da433
gkanno@54b9525

(Are you waiting now ?)

[diovisgood]

Thank you!
Though I had to migrate to PyTorch