Revisit when we define fit, transform, and fit_transform

[kwinkunks]

If code is in transform, it would also run during predict phase. But would only run if not last thing in pipeline... need to think about that.

I.e. if pipeline is [scaler, detector, estimator], then fit_transform is called in fitting, and transform is called in scoring or prediction. OTOH, if pipeline is [scaler, detector] then only fit is called in fitting, and predict is called in scoring or prediction (which these things are not designed for so don't implement -- but check this behaviour!).

Note that y is not passed to transform from fit_transform.

So detectors should work like:

• fit: call transform
• transform: put the detector here but note that y is not passed to transform from fit_transform
• fit_transform: call transform (i.e. skip fit) because otherwise y is not passed

Should only use fit when we want to remember something from training for comparison later.