pmelchior / proxmin

Proximal optimization in pure python
MIT License
113 stars 24 forks source link

Nesterov admm #9

Closed pmelchior closed 7 years ago

pmelchior commented 7 years ago

This PR provide Nesterov acceleration for PGM, and ADMM-derivates if they don't use a prox_g, i.e. if all proximal operations are equivalent to forward-backward.

The reason this does not work with prox_g is that the acceleration only applies to X, but there is no clear way to accelerate Z and U, or even if they should be accelerated.

Nonetheless, for the NMF we often only put projection-style constraints on each variables, and in this case acceleration is useful.