@glemaitre @betatim can you help me with the review? I just corrected the merge conflict in the changelog.

We discuss this PR IRL with @jeremiedbb and @ogrisel to know if we really wanted this feature.

In the end, the feature will be useful if we succeed to avoid some memory copy. Otherwise, we could instead have a FunctionTransformer whose job is to call astype before or after the scaler in a Pipeline.

So here, the fix is not the right one since we actually do a copy when calling _validate_data. Putting fit aside, at transform, we should keep X dtype untouched and instead pre-allocate X_trans to be returned in the output dtype. Then, all operations could be done using the out from numpy. In this case, we never modify X and thus avoid an additional memory copy.

Where it becomes more cumbersome is handling the sparse case since the API offered by scipy.sparse is not the same. It means that we will have diverging code between dense and sparse which is not the case for the moment.

This PR could be useful to actually evaluate how much cumbersome and additional maintenance effort is required if we want to go this path.

Closing this PR as we are not sure we want this feature (and would need different implementation) and @ducatte indicated they were not interested in continuing on this but please re-open if desired.

I've copied @glemaitre 's useful repy to the issue #18443