d-tiapkin/odcgm

The implementation for the paper "Orthogonal Directions Constrained Gradient Method: from non-linear equality constraints to Stiefel manifold"

Orthogonal Directions Constrained Gradient Method: from non-linear equality constraints to Stiefel manifold

The official implementation for the paper "Orthogonal Directions Constrained Gradient Method: from non-linear equality constraints to Stiefel manifold".

The presented algorithm generalized the Landing algorithm (see https://github.com/pierreablin/landing) from the manifold of orthogonal matrices to Stiefel manifold. The implementation is inpired by the code of Landing algorithm.

Use

The main algorithmic part is presented in the file algo.py. There is an pytorch optimizer ODCGM_SGD that mimics geoopt's RiemannianSGD.