Implement MetaLearnerGridSearch

[FrancescMartiEscofetQC]

This PR implements GridSearchCV which performs an exhaustive search over all the parameters combination. I restricted the search to only one type of metalearner because the model names are different for each metalearner and therefore there were many cases and the user can perform one search for each metalearner type and would be the same.

With the current setup model reusage does not work. This is because the grid search performs data splitting and therefore in-sample predictions for training evaluation can't be performed. One option would be to only compute test losses or not support model reusage.

Checklist

• Added a CHANGELOG.rst entry

[kklein]

With the current setup model reusage does not work. This is because the grid search performs data splitting and therefore in-sample predictions for training evaluation can't be performed. One option would be to only compute test losses or not support model reusage.

Two thoughts:

• If we allowed for the passing of an optional KFold parameter to MetaLearnerGridSearchCV.__init__ we could reuse the same splits across MetaLearnerGridSearchCV and thereby reuse models between MetaLearnerGridSearchCV s.
• If we had a evalute_in_sample_flag in MetaLearnerGridSearchCV.fit, a user could turn off the in-sample evaluation and thereby use any pretrained model (unless I'm missing something, of course).

What do you think of these options?

[FrancescMartiEscofetQC]

• If we allowed for the passing of an optional KFold parameter to MetaLearnerGridSearchCV.__init__ we could reuse the same splits across MetaLearnerGridSearchCV and thereby reuse models between MetaLearnerGridSearchCV s.

I think this could work but we would have the same problem as we had when synchronizing of needing to pass the indices as just with KFold if random_state was None it's not reproducible so I wouldn't suggest it.

• If we had a evalute_in_sample_flag in MetaLearnerGridSearchCV.fit, a user could turn off the in-sample evaluation and thereby use any pretrained model (unless I'm missing something, of course).

I think this is quite nice but the problem is at fit time for metalearners which have two stages the output of the predictions for the nuisance models are done with is_oos=False and therefore this raises an error because the fitting of the nuisance model was done in a different dataset.

[kklein]

Mostly small things :)

[kklein]

I'm afraid I don't quite see yet why we need/want the transformation from
{key: [value1, value2, value3]} to [{key: value1}, {key: value2}, {key: value3}] :/

[FrancescMartiEscofetQC]

We don't need it at the __init__ so I moved this conversion to the fit.
d6c8c3f

[kklein]

LGTM - thank you! :)