LinearPredictor.Statistics is always null

[klausmh]

System information

• OS version/distro: Win 10 / ML.NET 0.5.0
• .NET Version (eg., dotnet --info): 4.6.2

Issue

• What did you do?

I train a model using logistic regression:

var trainer = new LogisticRegression(
                        environment,
                        new LogisticRegression.Arguments() { ShowTrainingStats = true });
var trainRoles = ...
var predictor.Train(trainRoles)

• What happened?
  predictor.Statistics is null

• What did you expect?
  To be able to retrieve model statistics

Source code / logs

LogisticRegression.CreatePredictor does not pass its _stats LinearModelingStatistics field when it creates a LinearBinaryPredictor:

machinelearning/src/Microsoft.ML.StandardLearners/Standard/LogisticRegression/LogisticRegression.cs

Line 327 in 1e7b8be

new LinearBinaryPredictor(Host, ref weights, bias),

.

If I do change that, then predictor.Statistics.GetCoefficientStatistics still returns null. It appears that you calculate the inverted Hessian, but then do not use it to calculate standard errors (and pass these in the constructor of `LinearModelStatistics'):

machinelearning/src/Microsoft.ML.StandardLearners/Standard/LogisticRegression/LogisticRegression.cs

Line 189 in 1e7b8be

// Building the variance-covariance matrix for parameters.

.

[montebhoover]

So there are two separate issues here:

1. _stats is not passed to predictor: captured in issue Logistic Regression doesn't work with training stats #1205
2. GetCoefficientStatistics are no longer calculated since decoupling from MKL: captured in issue Training Statistics for linear learners. #612

[GalOshri]

@klausmh, would you be able to take a look at the discussion in #612 and the relevant PR (#1392) to make sure the solution will fulfill your requirements?

[sfilipi]

Closed with PR #1392