public static void SaveAll()
{
if (BatchStart != null)
{
BatchStart.Invoke(All, EventArgs.Empty);
}
All.ForEach(m => m.Save());
if (BatchEnd != null)
{
BatchEnd.Invoke(All, EventArgs.Empty);
}
}
private double TrainBatches(Matrix trainingData, Matrix labels, int batchSize, int epoch)
{
int currentIndex = 0;
int currentBatch = 1;
List <double> batchLosses = new List <double>();;
//Loop untill the data is exhauted for every batch selected
while (trainingData.CanSliceRows(currentIndex, batchSize))
{
//Get the batch data based on the specified batch size
var xtrain = trainingData.SliceRows(currentIndex, batchSize);
var ytrain = labels.SliceRows(currentIndex, batchSize);
//Run forward for all the layers to predict the value for the training set
var ypred = Forward(xtrain);
//Find the loss/cost value for the prediction wrt expected result
var costVal = Cost.Forward(ypred, ytrain);
batchLosses.Add(costVal.Data[0]);
//Get the gradient of the cost function which is then passed to the layers during back-propagation
var grad = Cost.Backward(ypred, ytrain);
//Run back-propagation accross all the layers
Backward(grad);
//Now time to update the neural network weights using the specified optimizer function
foreach (var layer in Layers)
{
Optimiser.Update(layer);
}
currentIndex = currentIndex + batchSize;
double batchLossAvg = Math.Round(costVal.Data[0], 3);
BatchEndEventArgs eventArgs1 = new BatchEndEventArgs(epoch, currentBatch, batchLossAvg);
BatchEnd?.Invoke(epoch, eventArgs1);
currentBatch += 1;
}
return(Math.Round(batchLosses.Average(), 3));
}
/// <summary>
/// Train the model with training dataset, for certain number of iterations and using batch size
/// </summary>
/// <param name="x"></param>
/// <param name="y"></param>
/// <param name="numIterations"></param>
/// <param name="batchSize"></param>
public void Train(NDArray x, NDArray y, int numIterations, int batchSize)
{
//Initialise bacch loss and metric list for temporary holding of result
List <double> batchLoss = new List <double>();
List <double> batchMetrics = new List <double>();
//Loop through till the end of specified iterations
for (int i = 1; i <= numIterations; i++)
{
//Initialize local variables
int currentIndex = 0;
batchLoss.Clear();
batchMetrics.Clear();
//Loop untill the data is exhauted for every batch selected
while (x.Next(currentIndex, batchSize))
{
//Get the batch data based on the specified batch size
var xtrain = x.Slice(currentIndex, batchSize);
var ytrain = y.Slice(currentIndex, batchSize);
//Run forward for all the layers to predict the value for the training set
var ypred = Forward(xtrain);
//Find the loss/cost value for the prediction wrt expected result
var costVal = Cost.Forward(ypred, ytrain);
batchLoss.AddRange(costVal.Data);
//Find the metric value for the prediction wrt expected result
if (Metric != null)
{
var metric = Metric.Calculate(ypred, ytrain);
batchMetrics.AddRange(metric.Data);
}
//Get the gradient of the cost function which is the passed to the layers during back-propagation
var grad = Cost.Backward(ypred, ytrain);
//Run back-propagation accross all the layers
Backward(grad);
//Now time to update the neural network weights using the specified optimizer function
foreach (var layer in Layers)
{
Optimizer.Update(i, layer);
}
currentIndex = currentIndex + batchSize;;
}
//Collect the result and fire the event
double batchLossAvg = Math.Round(batchLoss.Average(), 2);
double batchMetricAvg = Metric != null?Math.Round(batchMetrics.Average(), 2) : 0;
TrainingLoss.Add(batchLossAvg);
if (batchMetrics.Count > 0)
{
TrainingMetrics.Add(batchMetricAvg);
}
EpochEndEventArgs eventArgs = new EpochEndEventArgs(i, batchLossAvg, batchMetricAvg);
BatchEnd?.Invoke(i, eventArgs);
}
}
/// <summary>
/// Called when [batch end].
/// </summary>
/// <param name="epoch">The epoch.</param>
/// <param name="batch">The batch.</param>
/// <param name="loss">The loss.</param>
/// <param name="metric">The metric.</param>
protected void OnBatchEnd(int epoch, int batch, float loss, float metric)
{
BatchEnd?.Invoke(this, new BatchEndEventArgs(epoch, batch, loss, metric));
}
