gist of using the *CUDA modules

gistfile1.txt

features = nn.Sequential()
features:add(nn.Transpose({1,4},{1,3},{1,2}))
features:add(nn.SpatialConvolutionCUDA(fSize[1], fSize[2], 9, 9, 2, 2)) -- (111 - 9 + 2)/2 = 52
features:add(nn.Threshold(0,1e-6))
features:add(nn.SpatialMaxPoolingCUDA(2,2,2,2)) -- 26
features:add(nn.SpatialConvolutionCUDA(fSize[2], fSize[3], 5, 5)) -- 22
features:add(nn.Threshold(0,1e-6))
features:add(nn.SpatialMaxPoolingCUDA(2,2,2,2)) -- 11
features:add(nn.SpatialConvolutionCUDA(fSize[3], fSize[4], 4, 4)) -- 8
features:add(nn.Threshold(0,1e-6))
features:add(nn.SpatialConvolutionCUDA(fSize[4], fSize[5], 3, 3)) -- 6
features:add(nn.Threshold(0,1e-6))
features:add(nn.SpatialMaxPoolingCUDA(2,2,2,2)) -- 3
features:add(nn.Transpose({4,1},{4,2},{4,3}))
features:add(nn.Reshape(featuresOut))

Why do you use a Threshold module? Are you using it as a ReLU?

oops, gist doesn't have notifications on comments i think, didn't notice it until now.

Yes, ReLU is equivalent to nn.Threshold(0,0)