Example of multiclass classification with random forest?

[gilnahmias]

Looking to train multiclass classification with random forest and hyperparameter tuning.

Tried (on dotnetcore with either the latest LightGbm nuget or cloning the repo directly):

var pipeline = new LearningPipeline();

pipeline.Add(new TextLoader(_dataFilePath).CreateFrom<MyDataClass>());

 // for features
pipeline.Add(new TextFeaturizer("My_Feature1_Vectorized","My_Feature1"));
pipeline.Add(new TextFeaturizer("My_Feature2_Vectorized","My_Feature2"));

// for the label
pipeline.Add(new TextFeaturizer("Label_Vectorized", "Label"));

// ** CAN'T FIND FAST FOREST MULTI CLASS CLASSIFIER ** //
pipeline.Add(new LightGbmClassifier() { NumLeaves = 5, NumTrees = 5, MinDocumentsInLeafs = 2 });

var model = pipeline.Train<MyDataClass, MyPredictionClass>();

This code crashes with System.InvalidOperationException: Entry point 'Trainers.LightGbmClassifier' not found

Also, if I'm missing any crucial steps (say around conversion from numbers to text or vice versa), please point them out.

[sfilipi]

Thanks for reporting, gilnahmias. What version of ML.Net are you using?

[gilnahmias]

Nuget: Microsoft.ML.LightGBM 0.3.0

[codemzs]

Can you please do dotnet publish?

[gilnahmias]

Thanks for the prompt reply, folks! Just tried dotnet publish but the result hasn't changed :(

[codemzs]

How is your code even compiling? LightGBM does not have any parameter by name NumTrees, it’s called NumBoostRound.

[gilnahmias]

I'm sorry - you're right. The code does read:

pipeline.Add(new LightGbmClassifier() { NumLeaves = 5, NumBoostRound = 5, MinDataPerLeaf = 2 });

It was a copy error as I renamed classes to avoid IP leakage.

[codemzs]

https://docs.microsoft.com/en-us/dotnet/api/microsoft.ml.trainers.lightgbmclassifier?view=ml-dotnet#properties

[gilnahmias]

Exactly. The line I just posted was copied from this link (there's an example at the top). Can you spot any errors?

[codemzs]

Have you tried all of the below workarounds?

forcing the assembly to be loaded, like: new LightGbmArguments();
dotnet publish your application and running it from the published folder (since all dependencies are copied during publish)
Using full .NET Framework

[gilnahmias]

Forced load the assembly using your first suggestion. The following code runs for a few minutes but returns an empty metrics object from the evaluator.Evaluate() method. Still trying to find out why.

// Workaround to ensure assembly is loaded
new LightGbmArguments();

var pipeline = new LearningPipeline();

pipeline.Add(new TextLoader(_dataFilePath).CreateFrom<MyDataClass>());

 // for features
pipeline.Add(new TextFeaturizer("My_Feature1_Vectorized","My_Feature1"));
pipeline.Add(new TextFeaturizer("My_Feature2_Vectorized","My_Feature2"));

// for the label
pipeline.Add(new Dictionarizer("Label"));

// Puts all features into a vector
pipeline.Add(new ColumnConcatenator("Features", "My_Feature1_Vectorized", "My_Feature2_Vectorized"));

pipeline.Add(new LightGbmClassifier() { NumLeaves = 5, NumBoostRound = 5, MinDataPerLeaf = 2 });

// Convert the Label back into original text(after converting to number by the Dictionarizer)
pipeline.Add(new PredictedLabelColumnOriginalValueConverter() { PredictedLabelColumn = "PredictedLabel" });

// Train
var model = pipeline.Train<MyDataClass, MyPredictionClass>();

// Evaluate
var testData = new TextLoader(_testFile).CreateFrom<MyDataClass>();
var evaluator = new ClassificationEvaluator();
var metrics = evaluator.Evaluate(model, testData);

[codemzs]

Test dataset schema was different from train dataset.