Probability is missing from the prediction Output schema of BinaryClassification.Trainers.AveragedPerceptron

[bsoman3]

Issue

Using the Averaged Perceptron Binary Classifier in the Pipeline:

var pipeline =
//Other things in the pipeline
.Append(mlContext.BinaryClassification.Trainers.AveragedPerceptron(learningRate: lr, numIterations: 5));

//Fit Model steps
//Save Model steps
//Load Model steps

var predictions = loadedModel.Transform(data);
var metrics = mlContextTest.BinaryClassification.Evaluate(predictions);

Leads to the following error.

'Probability column 'Probability' not found
Parameter name: schema'

What Happened

Looking at the outputSchema of the predictions IDataView, the probability column is absent.

Expected Behavior

The Probability column should be available in the predictions IDataView based on reply by @zeahmed here- #376 (comment)

Looking at other binary classifiers like fast tree, that column is present in their output schema.

System information

• Product: dotnet-ml-api
• GitHub Login: @bsoman3
• Microsoft Alias: bhsoman

• OS version/distro:Windows 10 Enterprise
• .NET Version (eg., dotnet --info): 4.6.1

[wschin]

AveragedPerceptron doesn't have probability output due to its math properties. To get probability, you need to add a calibrator. Please see

machinelearning/docs/samples/Microsoft.ML.Samples/Dynamic/Trainers/BinaryClassification/Calibrators/Platt.cs

Line 42 in 29ca1f8

var calibratorEstimator = mlContext.BinaryClassification.Calibrators.Platt();

where it first trains AveragedPerceptron and then adds a calibrator.

[rogancarr]

You can also use your code as is, but evaluate with the EvaluateNonCalibrated evaluator.

mlContext.BinaryClassification.EvaluateNonCalibrated()

This is the preferred route if you don't actually need a probability or want probabilistic evaluation metrics.

[bsoman3]

Makes sense! thanks for the input! This can be closed afaik