Fixed inference results when cate_feature_names is not defined

[moprescu]

No description provided.

[heimengqi]

Looks good!

[kbattocchi]

This is great. But could you please also add a test that covers the else branch so that we don't regress this change in the future?

[moprescu]

This is great. But could you please also add a test that covers the else branch so that we don't regress this change in the future?

@kbattocchi I thought about that, but for that to work, we need a CATE estimator with linear final stage that will never have cate_feature_names defined. If you have a better idea for testing, let me know!

[kbattocchi]

This is great. But could you please also add a test that covers the else branch so that we don't regress this change in the future?

@kbattocchi I thought about that, but for that to work, we need a CATE estimator with linear final stage that will never have cate_feature_names defined. If you have a better idea for testing, let me know!

I think that one fairly straightforward approach would be to create a very simple wrapper like:

class NoFtNamesEst:
    def __init__(self, cate_est):
        self.cate_est = clone(cate_est, safe=False)

    def __getattr__(self, name):
        if name != 'cate_feature_names`:
            return getattr(self.cate_est, name)
        else:
            return self.__getattribute__(name)

and then test that you can get the inference results with either a plain DML estimator or a wrapped one.

[kbattocchi]

Consider renaming the test, but otherwise looks great.

[kbattocchi]

I'd rename to something like test_inference_no_feature_names or something.

[moprescu]

I named it this way so we can add more tests under the same method in the future. And this is just one subtest.

[kbattocchi]

Okay; I generally prefer to have a larger number of fine grained tests than a smaller number of broader tests where possible because it makes it quicker to track down the nature of the error when there's a failure, but I don't feel super strongly about it.