TileLayer

[jeffdonahue]

This adds a layer to duplicate an input Blob N times along any axis. AKA repmat. (This is less general than either np.tile or repmat as those allow arbitrary tiling along all axes, but this satisfies a lot of use cases and could be generalized later, or even now, chained together with other Tile layers to get the more general behavior in a less convenient and probably less efficient way.)

[jeffdonahue]

Oh yeah, as noted by @petergehler in #966, this can already be done by a combination of SplitLayer with N tops and ConcatLayer with the corresponding N bottoms and concat_param { axis: K } where you'd set tile_param { tiles: N axis: K } here. I'm not actually sure if this is more efficient than the combination of those layers, but I think it's a bit easier and more natural to write...

[myfavouritekk]

INSTANTIATE_LAYER_GPU_FUNCS(TileLayer); and INSTANTIATE_CLASS(TileLayer); seem to cause duplicate explicit instantiation of Forward_gpu and Backward_gpu. I put the else block in a new tile_layer.cu and solved the problem.

[jeffdonahue]

This is rebased and ready for review. I also added CUDA kernels for the GPU implementation in the latter commit.

[shelhamer]

@jeffdonahue this looks ready to me once it's rebased for the layer param ID since #2032 was merged. Feel free to update and merge.

[jeffdonahue]

Rebased. Thanks for reviewing @shelhamer!

[haihaoshen]

Hi, would you please take a look at TileLayer Usage #4422