Prior binary trainer requires float label column, but all other binary trainers require bool

[eerhardt]

Build a pipeline like:

            IEstimator<ITransformer> pipeline =
                MLContext.Transforms.ReplaceMissingValues("FixedAcidity", replacementMode: MissingValueReplacingEstimator.ReplacementMode.Mean)
                .Append(MLContext.FloatToBoolLabelNormalizer())
                .Append(MLContext.Transforms.Concatenate("Features",
                    new[]
                    {
                        "FixedAcidity",
                        "VolatileAcidity",
                        "CitricAcid",
                        "ResidualSugar",
                        "Chlorides",
                        "FreeSulfurDioxide",
                        "TotalSulfurDioxide",
                        "Density",
                        "Ph",
                        "Sulphates",
                        "Alcohol"}))
                .Append(MLContext.BinaryClassification.Trainers.Prior());

Scenario taken from XamlBrewer.Uwp.MachineLearningSample with the "Random" trainer changed to the "Prior" trainer.

When you try to fit this pipeline, you get an exception:

{System.ArgumentOutOfRangeException: Invalid type for Label column
Parameter name: data
 	Microsoft.ML.Core.dll!Microsoft.ML.Runtime.Contracts.CheckParam(Microsoft.ML.Runtime.IExceptionContext ctx, bool f, string paramName, string msg) Line 542	C#
 	Microsoft.ML.StandardTrainers.dll!Microsoft.ML.Trainers.PriorTrainer.Train(Microsoft.ML.TrainContext context) Line 235	C#
 	Microsoft.ML.StandardTrainers.dll!Microsoft.ML.Trainers.PriorTrainer.Microsoft.ML.ITrainer<Microsoft.ML.Trainers.PriorModelParameters>.Train(Microsoft.ML.TrainContext context) Line 274	C#
 	Microsoft.ML.StandardTrainers.dll!Microsoft.ML.Trainers.PriorTrainer.Fit(Microsoft.ML.IDataView input) Line 222	C#
 	Microsoft.ML.Data.dll!Microsoft.ML.Data.EstimatorChain<Microsoft.ML.ITransformer>.Fit(Microsoft.ML.IDataView input) Line 67	C#
>	XamlBrewer.Uwp.MachineLearningSample.exe!XamlBrewer.Uwp.MachineLearningSample.Models.BinaryClassificationModel.BuildAndTrain(string trainingDataPath, Microsoft.ML.IEstimator<Microsoft.ML.ITransformer> algorithm) Line 46	C#

It appears the Prior binary trainer can only take float labels:

machinelearning/src/Microsoft.ML.StandardTrainers/Standard/Simple/SimpleTrainers.cs

Line 233 in 45b6c43

_host.CheckParam(labelCol.Type == NumberDataViewType.Single, nameof(data), "Invalid type for Label column");

However, all other binary trainers require the label column to be a boolean.

cc @TomFinley

[codemzs]

@eerhardt FastTree seems to be recognizing more than just bool:

public static void CheckBinaryLabel(this RoleMappedData data)
{
Contracts.CheckValue(data, nameof(data));

        if (!data.Schema.Label.HasValue)
            throw Contracts.ExceptParam(nameof(data), "Training data must specify a label column.");
        var col = data.Schema.Label.Value;
        Contracts.Assert(!col.IsHidden);
        if (col.Type != BooleanDataViewType.Instance && col.Type != NumberDataViewType.Single && col.Type != NumberDataViewType.Double && !(col.Type is KeyDataViewType keyType && keyType.Count == 2))
        {
            KeyDataViewType colKeyType = col.Type as KeyDataViewType;
            if (colKeyType != null)
            {
                if (colKeyType.Count == 1)
                {
                    throw Contracts.ExceptParam(nameof(data),
                        "The label column '{0}' of the training data has only one class. Two classes are required for binary classification.",
                        col.Name);
                }
                else if (colKeyType.Count > 2)
                {
                    throw Contracts.ExceptParam(nameof(data),
                        "The label column '{0}' of the training data has more than two classes. Only two classes are allowed for binary classification.",
                        col.Name);
                }
            }
            throw Contracts.ExceptParam(nameof(data),
                "The label column '{0}' of the training data has a data type not suitable for binary classification: {1}. Type must be Boolean, Single, Double or Key with two classes.",
                col.Name, col.Type);
        }
    }

[TomFinley]

Interesting @codemzs. Some of the trainers don't use our "conveniences" for ingesting labels, due to them having unusual patterns of access of the data. FastTree in particular does a "transpose" of the data, so I can easily imagine this being the case, that it was missed, just as prior trainer was. (Prior trainer is likewise unusual, in that unlike other trainers it does not try to fetch a feature vector at all.) So we should make sure it's consistent. I wonder if LightGBM has a similar problem...

[eerhardt]

I thought @Ivanidzo4ka made all binary trainers only accept booleans with #2804.

However, looking through that PR I see #2804 (comment)