This repository contains an alternative implementation of SPDnet from:
- Davoudi
- torchspdnet
Main differences are:
- Extensive relying on
torch.einsum
for increased computational speed - Grad of BiMap coded with
einsum
rather than usingautograd
- Works for any number of batch dimensions (rather than 1 for Davoudi and 2 for torchspdnet)
- bias term on the eigenvalues in the reEig layers ?