Neuraxio / Neuraxle

The world's cleanest AutoML library ✨ - Do hyperparameter tuning with the right pipeline abstractions to write clean deep learning production pipelines. Let your pipeline steps have hyperparameter spaces. Design steps in your pipeline like components. Compatible with Scikit-Learn, TensorFlow, and most other libraries, frameworks and MLOps environments.
https://www.neuraxle.org/
Apache License 2.0
608 stars 62 forks source link

Feature: Remove narrow_space_from_best_guess function for HyperparameterSpace, and find alternative to nullify #447

Closed vincent-antaki closed 2 years ago

vincent-antaki commented 3 years ago

HyperparameterSpace has two function which aren't used, or tested, anywhere (nullify and narrow_space_from_best_guess). I believe they may be remnant of old hyperparameter research strategy. In my humble opinion, they should be removed. or at the very least moved to a class which implement the hyperparam research strategy interface.

guillaume-chevalier commented 3 years ago

Nice catch. narrow_space_from_best_guess should be removed without further questionning. Note that it is explained in some documentation example notebooks on the sphinx site, so the according documentation shall be removed as well upon doing this.

For the nullify method, I think that despite this method may seem available, that it may not work properly already. If done correctly, nullifying could be an indicator for the TPE of which Optional(step...) variables to ignore (or consider) for certain trials into how it accounted, or not, in the score obtained. Other hyperparameter selection strategies we would further code could make use of nullified params as well.

If we remove the idea of nullifying some params, we need a replacement for indicating which parameters are nullified from optionals. Perhaps a new apply method like pipeline.get_nullified_hyperparams() which only optionals would implement, and which would be useable by (or sent to) the TPE ?

@vincent-antaki

guillaume-chevalier commented 3 years ago

Related: #254