static void Main(string[] args)
{
var context = new MLContext();
Console.WriteLine("Loading Data...");
var colDef = new TextLoader.Column[] {
new TextLoader.Column(nameof(Digit.PixelValues), DataKind.Single, 1, 784),
new TextLoader.Column("Number", DataKind.Single, 0)
};
var trainDataView = context.Data.LoadFromTextFile(trainDataPath, colDef, hasHeader: true, separatorChar: ',');
var testDataView = context.Data.LoadFromTextFile(testDataPath, colDef, hasHeader: true, separatorChar: ',');
var pipeline = context.Transforms.Conversion.MapValueToKey(outputColumnName: "Label", inputColumnName: "Number", keyOrdinality: ValueToKeyMappingEstimator.KeyOrdinality.ByValue)
.Append(context.Transforms.Concatenate("Features", nameof(Digit.PixelValues)))
.AppendCacheCheckpoint(context)
.Append(context.MulticlassClassification.Trainers.OneVersusAll(context.BinaryClassification.Trainers.FastForest(), "Label"))
.Append(context.Transforms.Conversion.MapKeyToValue(outputColumnName: "Number", inputColumnName: "Label"));
Console.WriteLine("Training the model...");
var model = pipeline.Fit(trainDataView);
Console.WriteLine("Evaluating model...");
var predictions = model.Transform(testDataView);
var metrics = context.MulticlassClassification.Evaluate(predictions, labelColumnName: "Number", scoreColumnName: "Score");
// show evaluation metrics
Console.WriteLine($"Evaluation metrics");
Console.WriteLine($" MicroAccuracy: {metrics.MicroAccuracy:0.###}");
Console.WriteLine($" MacroAccuracy: {metrics.MacroAccuracy:0.###}");
Console.WriteLine($" LogLoss: {metrics.LogLoss:#.###}");
Console.WriteLine($" LogLossReduction: {metrics.LogLossReduction:#.###}");
Console.WriteLine();
var digits = context.Data.CreateEnumerable <Digit>(testDataView, false).ToArray();
var testDigits = new Digit[] {
digits[215], // 0
digits[202], // 1
digits[199], // 2
digits[200], // 3
digits[198], // 4
digits[207], // 5
digits[201], // 6
digits[220], // 7
digits[226], // 8
digits[235] // 9
};
var engine = context.Model.CreatePredictionEngine <Digit, DigitPrediction>(model);
var table = new BetterConsoleTables.Table(TableConfiguration.Unicode());
table.AddColumn("Digits");
for (var i = 0; i < 10; i++)
{
table.AddColumn($"P{i}");
}
for (var i = 0; i < testDigits.Length; i++)
{
var prediction = engine.Predict(testDigits[i]);
table.AddRow(
testDigits[i].Number,
prediction.Score[0].ToString("P2"),
prediction.Score[1].ToString("P2"),
prediction.Score[2].ToString("P2"),
prediction.Score[3].ToString("P2"),
prediction.Score[4].ToString("P2"),
prediction.Score[5].ToString("P2"),
prediction.Score[6].ToString("P2"),
prediction.Score[7].ToString("P2"),
prediction.Score[8].ToString("P2"),
prediction.Score[9].ToString("P2"));
}
// show results
Console.WriteLine(table.ToString());
}
static void Main(string[] args)
{
// Create machine learning context.
var context = new MLContext();
// Let's load model if it already exists, otherwise let's create it.
var modelPath = @"Assets\model.zip";
if (File.Exists(modelPath))
{
// Load existing model.
_model = context.Model.Load(modelPath, out DataViewSchema modelSchema);
}
else
{
// Upload training data set.
LoadTrainingDigits();
// Load training data set to ML context.
var trainingData = context.Data.LoadFromEnumerable(_trainingDigits.OrderBy(d => d.Number));
trainingData = context.Data.Cache(trainingData);
// Define the trainer.
// Different trainers give different accuracy of prediction, e.g.:
// - LbfgsMaximumEntropy gives better accuracy but lasts longer (92%, > 10 min);
// - SdcaMaximumEntropy gives worse accuracy, but lasts shorter (87%, < 1 min).
var pipeline = context.Transforms.Conversion.MapValueToKey(inputColumnName: "Number", outputColumnName: "Label")
.Append(context.Transforms.Concatenate("Features", "Pixels"))
.Append(context.MulticlassClassification.Trainers.SdcaMaximumEntropy(labelColumnName: "Label", featureColumnName: "Features"))
.Append(context.Transforms.Conversion.MapKeyToValue(inputColumnName: "PredictedLabel", outputColumnName: "PredictedNumber"));
// Train the model.
_model = pipeline.Fit(trainingData);
// Save the model.
context.Model.Save(_model, trainingData.Schema, modelPath);
}
#region Evaluate model uging test data set
// Upload test data set.
LoadTestDigits();
// Load test data set to ML context.
var testData = context.Data.LoadFromEnumerable(_testDigits);
testData = context.Data.Cache(testData);
// Perform prediction on test data.
var predictions = _model.Transform(testData);
// Evaluate the model using prediction on test data.
var metrics = context.MulticlassClassification.Evaluate(predictions);
Console.WriteLine("Model's evaluation metrics:");
Console.WriteLine($" MicroAccuracy: {metrics.MicroAccuracy:0.###}");
Console.WriteLine($" MacroAccuracy: {metrics.MacroAccuracy:0.###}");
Console.WriteLine($" LogLoss: {metrics.LogLoss:#.###}");
Console.WriteLine($" LogLossReduction: {metrics.LogLossReduction:#.###}");
Console.WriteLine();
#endregion
// Use predict engine to predict single values.
var predictionEngine = context.Model.CreatePredictionEngine <Digit, DigitPrediction>(_model);
Digit digit;
DigitPrediction prediction;
// Recognize image from img.bmp file.
Console.WriteLine("Recognizing image from img.bmp file...");
digit = new Digit(_recognizeFile);
digit.Draw();
prediction = predictionEngine.Predict(digit);
Console.WriteLine($"Predicted number: {prediction.PredictedNumber}");
for (var i = 0; i < prediction.Score.Length; i++)
{
Console.WriteLine($"[{i}]: {prediction.Score[i].ToString("P2")}");
}
Console.WriteLine("");
// Recognize random digit from test data set.
Console.WriteLine("Recognizing random image from test data set...");
var rand = new Random();
digit = _testDigits.Skip(rand.Next(_testDigits.Length)).First();
digit.Draw();
prediction = predictionEngine.Predict(digit);
Console.WriteLine($"Original number: {digit.Number}, predicted number: {prediction.PredictedNumber}");
for (var i = 0; i < prediction.Score.Length; i++)
{
Console.WriteLine($"[{i}]: {prediction.Score[i].ToString("P2")}");
}
Console.WriteLine("");
}
