The gfdl library currently leverages the sklearn Ridge estimator when the Moore-Penrose exact solve is not used. One potential issue is that we do not provide the user with a way to select the "true" solver used by Ridge, instead delegating that choice to Ridge itself via the default solver='auto'.
There are a few potential problems with this design:
- When performing a disciplined hyperparameter search over possible estimator settings in the
gfdl library, the inability to select the exact Ridge solver used may require us to cripple i.e., the range of network sizes we may explore (since some solvers may have higher space complexities than others, and we can't know which one will get selected, so we'd have to be overly conservative).
- The user or researcher may have good reasons to want to select the
solver--for example, comparison with a previously published result, etc.
The
gfdllibrary currently leverages thesklearnRidgeestimator when the Moore-Penrose exact solve is not used. One potential issue is that we do not provide the user with a way to select the "true" solver used byRidge, instead delegating that choice toRidgeitself via the defaultsolver='auto'.There are a few potential problems with this design:
gfdllibrary, the inability to select the exactRidgesolver used may require us to cripple i.e., the range of network sizes we may explore (since some solvers may have higher space complexities than others, and we can't know which one will get selected, so we'd have to be overly conservative).solver--for example, comparison with a previously published result, etc.