Deconvolution layer not working in python interface

[young-geng]

When using python NetSpec to generate network prototxt, layers.Deconvolution would not accept any parameters. It produces this error: AttributeError: 'LayerParameter' object has no attribute 'stride'

A quick look through the source code indicates that this is likely caused by the python interface getting the all the layer names by listing the fields of LayerParameter message in the protobuf. However, since deconv layers uses ConvolutionParameter, it does not have a DeconvolutionParameter message and hence won't be listed in the python interface.

I've also uploaded a piece of code that would trigger the problem, and you can find it here.

[shelhamer]

For these cases you explicitly include the relevant param in the net spec:

n.deconv = L.Deconvolution(n.input, 
    convolution_param=dict(num_output=21, kernel_size=64, stride=32))

See this example: https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/net.py#L58-L61

Thanks for including all the details in your post, but this is working as expected. See #3954 for an extension that makes Deconvolution inherit the convolution_param to work as you have suggested.

[seanbell]

@shelhamer maybe #3954 should be merged, if it doesn't break anything?