ManifoldParameter

[camilobrownpinilla]

Hi,
Great project, and thank you for releasing some nice code! Regarding the difference between manifold parameters and euclidean parameters, I am wondering why manifold parameters are only registered for the HypCLS module? Also, I am probably just not understanding something correctly, but do you mind elaborating on the suggested use case for HypCLS vs. nn.Linear? Thanks!

[marlin-codes]

Thank you for your interest! HypCLS defines prototypes as categories and computes negative distances with softmax to obtain probabilities. The prototypes are defined in hyperbolic space and optimized using a Riemannian optimizer. However, for nn.Linear layers, we optimize parameters that exist in Euclidean space, so we employ a Euclidean optimizer. You could certainly define a Euclidean prototype-based classifier and optimize its parameters using a Euclidean optimizer.