TPOTClassifier.set_params doesn't follow scikit-learn estimator API

[TomAugspurger]

Typically, calling old_estimator.set_params(new_param=foo) will copy all the other parameters from the old estimator, and only update new_param. Currently, TPOT will __init__ with the kwargs passed to new_param, and use the default or everything else.

Context of the issue

See http://scikit-learn.org/stable/developers/contributing.html#estimators and http://scikit-learn.org/stable/developers/contributing.html#get-params-and-set-params.

Process to reproduce the issue

clf = TPOTClassifier(random_state=0)
clf.set_params(n_jobs=-1).random_state
# output is None, the default random_state for TPOTClassifier

Expected result

>>> clf.set_params(n_jobs=-1).random_state
0

Possible fix

Hopefully just delete the custom set_params :) That may not be possible though, as it looks like there's some parameter handling logic in the __init__. That logic would need to be either moved to fit (and transform) or a helper method.

[rhiever]

@amueller, any thoughts on how this should be resolved? Is it generally recommended to avoid parameter handling logic in the __init__ of an estimator? If so, we could move said logic to fit/predict.

[amueller]

Traveling right now. But yeah, no validation or anything in init. Sent from phone. Please excuse spelling and brevity.

…

On Thu, Aug 9, 2018, 16:14 Randy Olson ***@***.***> wrote: @amueller <https://github.com/amueller>, any thoughts on how this should be resolved? Is it generally recommended to avoid parameter handling logic in the __init__ of an estimator? If so, we could move said logic to fit/predict. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#739 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAbcFg01NXfaLVTQRzmwY1tLa_KyEKUpks5uPEO9gaJpZM4V1wnf> .

[weixuanfu]

OK, thanks! I think moving those logic to fit/predict is more clear way to fix this issue. I will submit a PR for it.

[weixuanfu]

This issue should be fixed in TPOT 0.9.4. Please feel free to reopen this issue if you have any questions.