aswi2021tri-musketyri-gitlab/Server/ServerApp/Predictor/SdcaMaximumEntropyClassifier.cs @ 9a529f10

//
// Author: Roman Kalivoda
//

using System;
using System.Collections.Generic;
using System.Linq;
using log4net;
using Microsoft.ML;

namespace ServerApp.Predictor
{
    /// <summary>
    /// Implementation of the LightGBM classifier in ML.NET.
    /// </summary>
    class SdcaMaximumEntropyClassifier : AbstractClassificationPredictor
    {
        private static readonly ILog _log = LogManager.GetLogger(typeof(SdcaMaximumEntropyClassifier));

        /// <summary>
        /// Instantiates new <c>MLContext</c>.
        /// </summary>
        public SdcaMaximumEntropyClassifier()
        {
            this._mlContext = new MLContext();
        }

        /// <summary>
        /// Loads existing model from file
        /// </summary>
        /// <param name="filename">path to the model file.</param>
        public SdcaMaximumEntropyClassifier(string filename) : this()
        {
            DataViewSchema modelSchema;
            this._trainedModel = _mlContext.Model.Load(filename, out modelSchema);
            // TODO check if the loaded model has valid input and output schema
            this._predictionEngine = _mlContext.Model.CreatePredictionEngine<ModelInput, ModelOutput>(this._trainedModel);
        }

        public override void Fit(IEnumerable<ModelInput> trainInput)
        {
            this._trainingDataView = _mlContext.Data.LoadFromEnumerable(trainInput);
            var pipeline = _mlContext.Transforms.Conversion.MapValueToKey(nameof(ModelInput.Label))
                .Append(_mlContext.Transforms.Conversion.ConvertType(nameof(ModelInput.Hour)))
                .Append(_mlContext.Transforms.Concatenate("Features", 
                new[] { nameof(ModelInput.Temp), nameof(ModelInput.Rain), nameof(ModelInput.Wind), nameof(ModelInput.Hour) }))
                .Append(_mlContext.Transforms.NormalizeMeanVariance("Features", useCdf:false))
                .AppendCacheCheckpoint(_mlContext)
                .Append(_mlContext.MulticlassClassification.Trainers.SdcaMaximumEntropy())
                .Append(_mlContext.Transforms.Conversion.MapKeyToValue(nameof(ModelOutput.PredictedLabel)));

            var cvResults = _mlContext.MulticlassClassification.CrossValidate(this._trainingDataView, pipeline);
            _log.Debug("Cross-validated the trained model");
            this._trainedModel = cvResults.OrderByDescending(fold => fold.Metrics.MicroAccuracy).Select(fold => fold.Model).First();
            _log.Info($"Selected the model #{cvResults.OrderByDescending(fold => fold.Metrics.MicroAccuracy).Select(fold => fold.Fold).First()} as the best.");
            this._predictionEngine = _mlContext.Model.CreatePredictionEngine<ModelInput, ModelOutput>(this._trainedModel);

        }    
    }
}