/// <summary>Trains the current network on the given training data,
/// with the given learning rate using Stochastic gradient descent.
/// Covers the training data in mini-batches of the given size for each
/// epoch.</summary>
/// <param name="trainingData">The training data to train on.</param>
/// <param name="learningRate">The constant rate of learning overal all
/// epochs.</param>
/// <param name="nEpochs">The number of epochs to train over.</param>
/// <param name="miniBatchSize">The size of each mini-batch.</param>
public void Train(TrainingData trainingData,
double learningRate, int nEpochs, int miniBatchSize, bool log = false)
{
miniBatchSize = Math.Min(miniBatchSize, trainingData.SampleSize);
for (int i = 0; i < nEpochs; i++)
{
// Randomise the training data
Console.WriteLine("Epoch " + i + " of " + nEpochs + ": Shuffling data...");
trainingData.Shuffle(rng);
/* Pick out as many mini-batches as we can to cover al training
* data */
int nMiniBatches = trainingData.SampleSize / miniBatchSize;
TrainingData[] miniBatches = new TrainingData[nMiniBatches];
for (int j = 0; j < nMiniBatches; j++)
{
miniBatches[j] = trainingData.GetMiniBatch(
j * miniBatchSize, miniBatchSize);
}
for (int j = 0; j < nMiniBatches; j++)
{
TrainingData miniBatch = miniBatches[j];
if (log)
{
Console.WriteLine(
"Processing mini-batch " + j + " of " + nMiniBatches);
}
stepGradientDescent(miniBatch, learningRate);
}
}
}
/// <summary>Performs one step of gradient descent according to the
/// current network's cost after training on the given mini-batch of
/// training data. Updates all weights and biases according to the
/// gradient of the cost function in the network's current
/// configuration.</summary>
/// <param name="miniBatch">The mini-batch of training data to use for
/// determining the cost of the current network.</param>
/// <param name="learningRate">The gradient descent step size.</param>
private void stepGradientDescent(TrainingData miniBatch, double learningRate)
{
int miniBatchSize = miniBatch.SampleSize;
Matrix[] weightsDirection = new Matrix[LayerCount - 1];
Matrix[] biasesDirection = new Matrix[LayerCount - 1];
/* Initialise the direction matrices to be of the same dimensions
* as the weights and biases */
for (int i = 0; i < LayerCount - 1; i++)
{
weightsDirection[i] = new Matrix(weights[i].RowCount,
weights[i].ColumnCount);
biasesDirection[i] = new Matrix(biases[i].RowCount, 1);
}
/* Sum the steps of weights and biases for each training example in
* the mini-batch */
for (int i = 0; i < miniBatchSize; i++)
{
Matrix input = Matrix.FromArray(miniBatch.GetInput(i));
Matrix expectedOutput = Matrix.FromArray(
miniBatch.GetExpectedOutput(i));
Tuple <Matrix[], Matrix[]> direction = getCostGradient(
input, expectedOutput);
for (int l = 0; l < LayerCount - 1; l++)
{
weightsDirection[l] += direction.Item1[l];
biasesDirection[l] += direction.Item2[l];
}
}
for (int l = 0; l < LayerCount - 1; l++)
{
weights[l] -=
(learningRate / miniBatchSize) * weightsDirection[l];
biases[l] -=
(learningRate / miniBatchSize) * biasesDirection[l];
}
}
