/// <summary>
/// Configure the pipeline settings.
/// </summary>
/// <param name="dataLoader">The data loader.</param>
private void ConfigurePipeline(ILearningPipelineDataLoader dataLoader)
{
var provider = (BatchDataProvider)dataLoader;
this.SetBatchSize(provider.BatchSize);
int datasetSize = DetermineDatasetSize(provider.LabelFilename);
this.SetDatasetSize(datasetSize);
}
/// <summary>
/// Runs the learning pipeline and generates a <see cref="PredictionModel"/>.
/// </summary>
/// <returns>The prediction model.</returns>
public PredictionModel Run()
{
ILearningPipelineDataLoader dataLoader = null;
ILearningPipelineOptimizer optimizer = null;
ILearningPipelineNeuralNetworkModel neuralNetworkModel = null;
foreach (ILearningPipelineItem item in this)
{
switch (item)
{
case ILearningPipelineDataLoader loader:
dataLoader = loader;
break;
case ILearningPipelineOptimizer opt:
optimizer = opt;
break;
case ILearningPipelineNeuralNetworkModel networkModel:
neuralNetworkModel = networkModel;
break;
}
}
if (dataLoader == null)
{
throw new ArgumentNullException(nameof(dataLoader));
}
if (optimizer == null)
{
throw new ArgumentNullException(nameof(optimizer));
}
if (neuralNetworkModel == null)
{
throw new ArgumentNullException(nameof(neuralNetworkModel));
}
// Indicate the start of training.
PipelineSettings.IsPipelingRunning = true;
PipelineSettings.CurrentIteration = 0;
for (var epoch = 0; epoch < PipelineSettings.EpochCount; epoch++)
{
PipelineSettings.CurrentEpoch = epoch + 1;
ConsoleUtility.WriteLine($"Current epoch: {PipelineSettings.CurrentEpoch}");
if (PipelineSettings.UseLearningRateDecay)
{
ApplyLearningRateScheduler(neuralNetworkModel);
}
for (var i = 0; i < PipelineSettings.TrainingIterationsCount; i++)
{
PipelineSettings.CurrentIteration++;
if (PipelineSettings.CanPerformDropout)
{
PipelineSettings.DropoutVectors = PipelineSettings.Dropout.GenerateDropoutVectors(PipelineSettings.HiddenLayerSizes);
}
// A training iteration is constited of three steeps:
// 1. Load data
var data = (MnistImageBatch)dataLoader.LoadData();
// 2. Feedforward step
double[][] prediction = neuralNetworkModel.Predict(data.Pixels);
// 3. Backpropagation step
optimizer.Optimize(prediction, data.Labels);
}
}
// Indicate the end of training.
PipelineSettings.IsPipelingRunning = false;
return(new PredictionModel((INeuralNetwork)neuralNetworkModel));
}
