Awesome

#extracunn

This package contains cuda implementations for various layers:

• SpatialConvolutionNoBias: convolutional layer without bias, useful for e.g. gated operations in spatial LSTM cells. <i>Note that currently the SpatialConvolution layer in torch has the option noBias(), which can be used instead of this layer.</i>
• Huber penalty: penalise for non-smoothness.
• MSSECriterion: scale-invariant loss layer, useful for e.g. depth estimation. Implementation from Eigen's paper. Note that the error is computed between log ground truth and log prediction.
• InterleaveTable: interleave elements of n tables, each one containing length elements.
• SplitTableMultiple: split a tensor into tensors of k elements each; for k=1, this is the same as SplitTable.
• SpatialSkew: skew input feature maps to the right; each row is shifted by one position wrt previous row.
• SpatialUnskew: undo the effect of SpatialSkew.
• SpatialMirrorHorizontal: mirror input tensor along central vertical axis.
• SpatialMirrorVertical: mirror input tensor along central horizontal axis.
• SpatialMirrorDiagonal: SpatialMirrorHorizontal + SpatialMirrorVertical done in one step for efficiency.

####Installation

Type 'luarocks make' inside the directory to install this package.