Dilated-Convolution-with-Learnable-Spacings-PyTorch

[0.1.1] - 2023-12-22

Added

• Added Dcls2dK1d class that is suitable for flat kernels in Conv2d ((7,1) kernel size for example. This method avoid extra computation for positions and sigmas along the flat dimension.

[0.1.0] - 2023-10-19

Fixed

• Fixed a clamp in the init of positions. This will not affect previous results as the clamp between -lim and lim for positions initialized with a centered normal law of std 0.5 has a very low probability of being activated given the large size of the kernels.

Added

• Added DCLSN_Md class for kernels where position learning is restricted to a subset of dimensions chosen from N kernel dimensions.
• Added ConstructKernel3_1d class, which constructs the appropriate kernel from the latter.
• Added Dcls3_1d class that does the appropriate convolution.

Changed

• DCLS default interpolation version is now gauss instead of v1 for all dimensions.