There are several solvers in cuML which can benefit algorithm development across the larger RAPIDS ecosystem by living in RAFT. For example, there is interest in exposing some of the solvers from RAFT to the cupy.optimize package.
Current solvers in cuml which we have discussed moving over to RAFT:
- L-BFGS (also to extend w/ trust region constraints for L-BFGS-B)
- OWL-QN
- Coordinate descent
- SGD
- LARS
- SMO
- Any solvers which will be exposed through the
cupy.optimize package will need to accept a generalized objective function (eg functor/lambda or constructed from some set of predefined building blocks like the semirings/distance API) and will need to accept constraints.
There have also been discussions of wanting the CG, Powell, and potentially Nelder-Meade simplex methods, though these would be targeted as future features after the above solvers have been moved over.
There are several solvers in cuML which can benefit algorithm development across the larger RAPIDS ecosystem by living in RAFT. For example, there is interest in exposing some of the solvers from RAFT to the
cupy.optimizepackage.Current solvers in cuml which we have discussed moving over to RAFT:
cupy.optimizepackage will need to accept a generalized objective function (eg functor/lambda or constructed from some set of predefined building blocks like the semirings/distance API) and will need to accept constraints.There have also been discussions of wanting the CG, Powell, and potentially Nelder-Meade simplex methods, though these would be targeted as future features after the above solvers have been moved over.