Pytorch implementation of the Sharpened Cosine Similarity layer mentioned in this tweet.
The current version supports square kernels with arbitrary kernel sizes.
Part of the implementation borrows from dvisockas/cos_sim.
- depth-wise conv (with groups arguement)
- non-square kernels
- efficiency
- demo training code