rescale regressors after acquisition

[lorenzozanisi]

I realised that the internal scaler of the regressors were not updated after the acquisition, so probably all the results so far are pessimistic; this should improve things. This might also explain why the regressor is worse when trained only on unstable points as opposed to pre-training it on both.

[lorenzozanisi]

The scaler is initialised when the Regressors are initialised, but then the data changes, and when you call unscale=True in the predict method the Regressor has no way of knowing that the data has changed and so it will produce rubbish unscaled results; unless you force the scaler inside the regressor to also change accordingly

[zcapjdb]

Ah I see but this would then only affect the unscaled losses in the output dictionary. The data is still correctly rescaled at each iteration and the results I showed were for the scaled losses so they would be unaffected by this change. So this doesn't explain the improved performance when including unstable points

[lorenzozanisi]

I think you're right. But it makes the predictions for the actual fluxes better, so in general it's a good thing to have.