Recommended way to handle "one parameter greater than the other" with _validate_params()?

[j-adamczyk]

I am writing a package compatible with Scikit-learn. I want to use the _validate_params() approach: #23462. Setting constraints for individual attributes is simple, and it works.

But I have min_distance and max_distance parameters, and require min_distance <= max_distance. Can this be achieved with _parameter_constraints dict? And if not, how should I approach this? I don't want to repeat validation code in .fit() and .transform(), but my transformer does not need any fitting (I have empty .fit() method).

I was thinking about overriding the _validate_params() in inheriting class and making the validation there. Will this suffice? Is this the good approach, or should I do something else?

[glemaitre]

But I have min_distance and max_distance parameters, and require min_distance <= max_distance. Can this be achieved with _parameter_constraints dict?

Nop it cannot.

I don't want to repeat validation code in .fit() and .transform(), but my transformer does not need any fitting (I have empty .fit() method).

A couple of things that might answer to your question:

• the parameter validation only happen at fit by decorating the method with _fit_context;
• we do not validate parameter at transform;
• what you describe looks like a stateless transformer (that follows as the above rule). You can check Normalizer implementation for an example.
• the constraint between parameters will be in fit.
• if the user does not call fit then he will be on his own regarding the validation.

[j-adamczyk]

Thanks, that was incredibly helpful. I think that for my use case I'll stick with calling _validate_params explicitly both in fit and transform, since I have a lot of stateless transformers.

Just one question though - why would the constraint be in fit? I think that overriding _validate_params and putting additional validation there is cleaner. This way, entire validation logic is in one function, either implicitly or explicitly.