Slice layer

[kSkip]

I am training a U-Net using the architecture from U-Net: Convolutional Networks for Biomedical Image Segmentation. It requires cropping the tensors from the encoding stages to concatenate with the outputs of the decoding stages. I created the slice layer to implement this.

AFAICK, there are no layers in dlib that accommodate this. However, I probably should have asked before creating this one.

Some of my considerations:

• It is similar to extract, but instead it copies non-contiguous regions to perform a 3D crop. This is both different behavior and performance.
• I could have used an alias_tensor, similarly to extract, by aliasing every row. The CPU implementation may have been similar performance doing this. However, this is not ideal with CUDA because it requires a kernel invocation for every row. Since kernels on the same stream execute serially, the GPU throughput would not be saturated. So, I overloaded copy_tensor to copy non-contiguous rows with a single kernel call.
• I don't know if the public interface is the best, but it works for my case. I'm happy to hear any suggestions or improvements so this can be useful to others.

Thanks for taking a look!

[davisking]

Nice yeah this seems like a useful thing to have :D

[davisking]

Add docs for this (like you see for the other functions in this file :) )

[kSkip]

Oh, I forgot that. Added the doc string. Thanks!

[davisking]

Thanks for the PR :)