Neural network library written in C#.
- Implementation of MLP neural network
- Gradient descent (with momentum) and Levenberg-Marquardt algorithms
- Various normalization algorithms
- Support of CSV files as separate package (high extensibility)
// load and divide data from CSV file
var csv = CsvFacade.LoadSets(@"C:\Users\Marek\Desktop\sin.csv",
new RandomDataSetDivider(),
new DataSetDivisionOptions { TrainingSetPercent = 80, ValidationSetPercent = 10, TestSetPercent = 10, });
// perform min-max data normalization
var normalization = new MinMaxNormalization(-1, 1);
await normalization.FitAndTransform(csv.sets);
// create 1 x 5 x 1 MLP neural network
var mlp = MLPNetwork.Create(
1, (5, new SigmoidActivationFunction()), (1, new LinearActivationFunction())
);
// use gradient descent algorithm
var algorithm = new GradientDescentAlgorithm(new GradientDescentParams
{
LearningRate = 0.001,
Randomize = true,
Momentum = 0.01,
});
// use quadratic loss function
var lossFunction = new QuadraticLossFunction();
// train neural network until error > 0.001
var trainer = new MLPTrainer(mlp, csv.sets, algorithm, lossFunction);
while (trainer.Error > 0.001)
{
trainer.DoEpoch();
Console.WriteLine($"Error: {trainer.Error}");
Console.WriteLine($"Validation error: {trainer.RunValidation()}");
}