Once the scikit-learn compatibility effort is done, it would be nice to refactor the learners a bit. Some things that could be done:
factor out all "stateful initialization" such as initializing random state, optimizer, regularizer and building the tf graph into a single function. Currently its spread around and called multiple times.
remove set_learnable_parameters. We get set_parameters for free from scikit-learn and all the complexity is gone once we've clearly contained the statefulness.
Once the scikit-learn compatibility effort is done, it would be nice to refactor the learners a bit. Some things that could be done:
set_learnable_parameters. We getset_parametersfor free fromscikit-learnand all the complexity is gone once we've clearly contained the statefulness.