To run it in fully Managed C#: Explanation on how to use the code. This information should be added to the repository

[78Spinoza]

Hello there First many thanks for the source code.
You can train your model in phyton or R , whatever and save it in native format. To run it in fully Managed C#: Explanation on how to use the code should be added to the repository Below is what I did to make it work and some small issues that needed to be resolved to get best performance.

Using LightGBM with C# in Fully Managed Code

1. Install the LightGBMNet.Train Package

• Use NuGet Package Manager or .NET CLI:

  dotnet add package LightGBMNet.Train

2. Load a Native LightGBM Model

• Set culture to InvariantCulture to avoid parsing errors.
• Use Booster.FromFile to load the model:

  Thread.CurrentThread.CurrentCulture = CultureInfo.InvariantCulture;
  var booster = Booster.FromFile(loadDataDialog.FileName);

3. Validate the Model

• Create a feature array and compare predictions with expected results:

  float[] inputData = new float[] { /* feature values */ };
  double shouldbe = -0.3346386538390942;
  double[] predictions = booster.PredictForMat(Booster.PredictType.Normal, inputData, 0, -1);
  double prediction = predictions[0];

4. Transform the Model to Fully Managed Code

• Use booster.GetModel() to get the managed model:

  (Ensemble ensemble, Parameters par) = booster.GetModel();

5. Validate the Managed Model

• Make predictions using the managed model and check for small rounding errors:

  VBuffer<float> denseBuffer = new VBuffer<float>(inputData.Length, inputData);
  ensemble.MaxThreads = Environment.ProcessorCount;
  double prediction2 = ensemble.GetOutput(ref denseBuffer, 0, -1);
  //   prediction2 = -0.33463865383909XX; where the  last digits XX where bad but this is ok for us. 

6. Save and Load the Managed Model

• Save the managed model using BinaryWriter and GZipStream:

  using (var fileStream = new FileStream(filePath, FileMode.Create, FileAccess.Write))
  using (var gzipStream = new GZipStream(fileStream, CompressionMode.Compress))
  using (var writer = new BinaryWriter(gzipStream))
  {
    ensemble.Save(writer);
  }

• Load the managed model:

  using (var fileStream = new FileStream(filePath, FileMode.Open, FileAccess.Read))
  using (var gzipStream = new GZipStream(fileStream, CompressionMode.Decompress))
  using (var reader = new BinaryReader(gzipStream))
  {
    loadedModel = new Ensemble(reader, false);
  }

7. Use the Managed Model in Production

• Use the managed model to avoid native DLL dependencies and run it on other OS like Linux.

By following these steps, you can effectively use LightGBM with C# in a fully managed environment, ensuring compatibility and performance across different platforms.

[mjmckp]

Thanks @78Spinoza, however these instructions are not how the library is intended to be used. See the unit tests to see how the native GBDTs can be trained, converted to managed objects, and evaluated. I do not recommend using the output of Booster.GetModel, as the ensemble object does not have all the necessary transformations on the output required for binary/multiclass model evaluation.

[78Spinoza]

I checked the unittest. I would like to train the model in Phyton since there are man many visualization and hyperparameter tuning that exist. How can I use a trained model and not Booster.GetModel ? I know that it only works for regression but internally all classification and other are regression for LightGBM but I understand what you say.. Can you please provide some example more simple as I did above?

[mjmckp]

@78Spinoza I'll have a look at how best to do this and let you know.