crop size and difference between cudnn and cuda-convnet about theano_alexnet

[gwding]

@soheilb could you paste the error message here? and which lines you were making changes?

[soheilb]

My changes are in the lines 37 and 45 of the alex_net.py file where I changed 227 to 224 (to make it exactly equal to the alexnet paper). Then I used the config file to select between cudnn and cuda-convnet (lib_conv flag). The error seems to be relevant to the size of the input for the first fully connected layer. Apparently using cudnn, the input size of the first fully connected layer changes to 6400 (256 times 25) while it is supposed to have dimension 9216 (256 times 36). Here is the error:

ValueError: dimension mismatch in args to gemm (256,6400)x(9216,4096)->(256,4096)
Apply node that caused the error: GpuDot22(GpuFlatten{2}.0, <CudaNdarrayType(float32, matrix)>)
Toposort index: 314
Inputs types: [CudaNdarrayType(float32, matrix), CudaNdarrayType(float32, matrix)]
Inputs shapes: [(256, 6400), (9216, 4096)]
Inputs strides: [(6400, 1), (4096, 1)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[GpuElemwise{Add}[(0, 0)](GpuDot22.0, GpuDimShuffle{x,0}.0)]]

[gwding]

@soheilb I would suggest looking at the padding size settings around
https://github.com/uoguelph-mlrg/theano_alexnet/blob/master/lib/layers.py#L172
There might be some subtle difference between the cudnn library and cudaconvnet, on how to use those settings.
Sorry that I personally don't have easy access on machines that can run these things, so couldn't try it myself.

[soheilb]

Yes, it seems the difference is in how they handle the padding. Thanks for looking into this.