/// <summary> Using the given <paramref name="trainingSet"/>, trains the neural network with the given <paramref name="batchSize"/> over the given <paramref name="epochs"/>. </summary>
/// <param name="random"> The random number generator used to shuffle the <paramref name="trainingSet"/>. </param>
/// <param name="trainingSet"> The <see cref="IDataSet"/> against which to train. </param>
/// <param name="epochs"> The number of times to go over the <paramref name="trainingSet"/>. </param>
/// <param name="batchSize"> The number of <see cref="IDataPoint"/>s to process before applying learning. </param>
/// <param name="logger"> <see cref="INetworkLogger"/> used to log the progress of the training. If this is <c>null</c>, no logging will happen. </param>
/// <param name="testSet"> Another <see cref="IDataSet"/>, with <see cref="IDataPoint"/>s that are not in the <paramref name="trainingSet"/>. If this is given, the network will be tested against the <paramref name="testSet"/> after every epoch. </param>
public void LearnOverData(Random random, IDataSet trainingSet, int epochs, int batchSize, INetworkLogger logger = null, IDataSet testSet = null)
{
// Create the various required arrays once, then just fill them with data. This cuts down on memory allocations, dramatically speeding up the training.
float[] input = new float[inputLayer.Count];
float[] output = new float[outputLayer.Count];
float[] desiredOutput = new float[outputLayer.Count];
float[] error = new float[outputLayer.Count];
// Calculate how many batches are required with the given batch size.
int numberOfBatches = trainingSet.Count / batchSize;
// Create the change to be made to the network.
NetworkChange networkChange = new NetworkChange(this);
// Do the given amount of epochs.
for (int currentEpoch = 0; currentEpoch < epochs; currentEpoch++)
{
// For each epoch, shuffle the training data randomly.
trainingSet.Shuffle(random);
// The percentage of data correctly guessed per batch.
float lastBatchPercentage = 0.0f;
// Do the given amount of batches.
for (int batch = 0; batch < numberOfBatches; batch++)
{
// How many data were correctly guessed in this batch.
int batchGuesses = 0;
// Do the calculated number of training data in the batch.
for (int dataIndex = 0; dataIndex < batchSize; dataIndex++)
{
// Get the next data and set the input array to its data.
IDataPoint currentData = trainingSet[batch * batchSize + dataIndex];
currentData.GetFloatData(ref input);
// Process the data through the network, saving its output.
ProcessInput(input, ref output);
// Get the output that the network guessed with the most confidence.
int bestGuess = output.GetIndexOfHighestValue();
// Create the desired output.
desiredOutput.ApplyVectorisedFunction((int g) => g == currentData.Label ? 1 : 0);
// If the guess was correct, track it.
if (bestGuess == currentData.Label)
{
batchGuesses++;
}
// Calculate the error.
CalculateError(desiredOutput, output, ref error);
// Learn from the output compared to the desired output, save the results in the previous network change.
CalculateChanges(networkChange, error);
// If a logger has been given, log the state of the network as of this iteration.
logger?.Log(currentEpoch, epochs, batch, numberOfBatches, dataIndex, batchSize, lastBatchPercentage);
}
// Calculate how many of the data in the batch were correctly guessed.
lastBatchPercentage = (float)batchGuesses / batchSize;
// Apply the changes.
ApplyChanges(networkChange, batchSize);
networkChange.Reset();
}
// If a logger was given and should log output, log the output of the epoch.
if (logger != null && (logger.OutputOptions & OutputOptions.WhenFinished) == OutputOptions.WhenFinished)
{
// Log that the epoch finished.
logger.Log($"\nEpoch {currentEpoch} finished");
// If test data was given, test against it and log the output.
if (trainingSet != null)
{
logger.Log($", correctly guessed {PercentageOfCorrectTestData(random, testSet, ref input, ref output):P2} of the test data");
}
// Log the full stop and line end.
logger.Log(".\n");
}
}
// If a logger was given, log that the training finished.
logger?.Log("Training finished.");
}
