Microsoft.ML nuget package no longer has a way to specify number of bins for binning normalization

[eerhardt]

In v0.11.0, it was possible to write code like the following:

var normalizer = mLContext.Transforms.Normalize(
    new NormalizingEstimator.BinningColumnOptions(outputColumnName: "Label", numBins: 2));

This allowed you to create a binning normalizer with the number of bins set to 2.

However, this API is no longer public in Microsoft.ML. There is a public API where you can specify mode: NormalizingEstimator.NormalizationMode.Binning, but there is no way to set the number of bins. So when you use this mode, you always MaximumBinCount set to the default 1024.

[codemzs]

@eerhardt You set max bins via experimental nuget, more here

[eerhardt]

I don't understand why I would have to use an experimental nuget package in order to use binning normalization. Especially since in the "stable" API there exists a Binning mode, but that mode is virtually useless if I can't specify how many bins I want.

[codemzs]

I'll let @artidoro comment on why this change was made.

[TomFinley]

At first glance this seems like an oversight we should correct. It seems like if you want nothing else out of your bin normalizer, you'd want to configure the number of discretization points. Add back in, maybe write somethign in the functional tests to cover it?

[TomFinley]

It seems like this is insufficient:

machinelearning/src/Microsoft.ML.Transforms/NormalizerCatalog.cs

Lines 26 to 29 in 3663320

	public static NormalizingEstimator Normalize(this TransformsCatalog catalog,
	string outputColumnName, string inputColumnName = null,
	NormalizingEstimator.NormalizationMode mode = NormalizingEstimator.NormalizationMode.MinMax)
	=> new NormalizingEstimator(CatalogUtils.GetEnvironment(catalog), outputColumnName, inputColumnName ?? outputColumnName, mode);

The way we solved this from the POV of the command line is we had minmax, GCN, and bin look like separate transforms. Maybe we ought to reflect that through this here. So: maybe there should be one method for minmax, another for GCN, another for bin, thereby allowing more detailed configuration since that's often useful.

While we're at it, I see we don't have the fix-zero configuration option settable, just an FYI, and that's pretty important.

[codemzs]

@TomFinley What do you feel about this?

[TomFinley]

@TomFinley What do you feel about this?

Looks good at first glance. Might need some input-output column pair overloads for the multicolumn mapping, but I don't insist on it. The people I view as having the most relevant feedback on this though are @artidoro (since he introduced it I believe) and @eerhardt (since he's the one that raised this issue).