review re-scaling code in estimator.predict()

[SteveOv]

The code which rescales the predictions in the estimator predict() function looks like;

        # Undo any scaling applied to the labels (e.g. the model predicts inc/100)
        if unscale:
            raw_preds /= [*self._labels_and_scales.values()]

Given that the raw_preds is a 3-d array of size (iters, insts, predictions) I'm concerned that this is not exlicitly told how to apply the scaling (effectively over the last axis). I think what's happening is that the axis is chosen because it matches the size of _labels_and_scales. My worry is that this may run into problems if the number of instances or the number of iterations matches the number of _labels_and_scales.

[SteveOv]

I've tested this and it's not a problem.

Looking at this more closely, I can see that the numpy broadcasting algorithm will work from the last to first dimension, to match the two arrays.

What is happening is exactly what needs to happen.

• the last dimension of the predictions is sized by the number of labels
• we are applying a 1d scaling list/array which is the number of labels long
• they match in length, so the scaling array is broadcast over the other two dimensions and the scaling is applied as required

For more info, see https://numpy.org/doc/stable/user/basics.broadcasting.html