MNT Param validation: Constraint representing objects exposing a given interface #23791
Conversation
|
#23621 should be merged first, so that we can update the constraint of the |
|
#23621 has been merged. So we can go forward with this PR. |
sklearn/utils/_param_validation.py
Outdated
| @validate_params({"method": [str]}) | ||
| def __init__(self, method): |
There was a problem hiding this comment.
In the future, I think we will need to extend this for two methods. For example, MultiOutputRegressor requires two methods: fit and predict:
scikit-learn/sklearn/multioutput.py
Lines 255 to 256 in 3c75d36
There was a problem hiding this comment.
Basically, I am wondering if we should not create a protocol for the minimal classifier, regressor and transformer.
There was a problem hiding this comment.
Would it be worth treating HasMethod similar to StrOptions in that you can provide a set of methods to check against (and calling it HasMethods). Then Is{Classifier,Regressor,Transformer,Estimator} could be special cases of that.
There was a problem hiding this comment.
Would it be worth treating HasMethod similar to StrOptions
I think so yes, with a slight difference from StrOptions, since the object would need to implement all listed methods and not one of the listed methods.
Basically, I am wondering if we should not create a protocol for the minimal classifier, regressor and transformer.
It would be interesting to explore, but I guess we can't start with what's in this PR first explore that later.
|
@jeremiedbb There is a conflict to resolve from #23470 |
This PR proposes to add a new constraint for the validation of parameters (#23462) to represent objects exposing a specific method.
This is useful for instance in
AgglomerativeClusteringhas amemoryparameter that accpets objects with acachemethod.LinearDiscriminantAnalysishas acovariance_estimatorparameter that expects an object with afitmethod.cc/ @glemaitre