Code cleanup and added pip installer

This patch brings reformatted code and a new notebook to compare hexagonal and square kernels. Furthermore, a pip installer is now available for convenience.

New functionalities and changed API

This release of HexagDLy brings new features like 3D convolutions and custom hexagonal kernels but also has a slightly changed API to be more PyTorch-like.

Changes

Dependencies

This version of HexagDLy requires Python >= 3.5 and was tested to work with PyTorch v0.4.1.

Syntax

The syntax has changed to follow torch.nn with capitals in the class names. A convolution e.g. is now defined with hexconv = hexagdly.Conv2d().

3D Convolutions

3 dimensional convolutions can be performed via hexagdly.Conv3d() where the hexagonal plane is indexed with the last two indices of the input tensor.

Custom Kernels

With the implemented hexagdly.Conv#d_CustomKernel() hexagonal kernels of 2 and 3 dimensions can be filled with arbitrary values via numpy arrays for each sub-kernel.

Initial release of HexagDLy

Initial release of HexagDLy