public void SGD(ref MnistData data, int epochs, int batchSize, float learningRate, bool output)
{
MathNet.Numerics.Control.UseSingleThread(); // Single thread is optimal over multithreading.
Console.WriteLine("Training...");
for (int l = 0; l < vSizes.Count; l++)
{
mSumW[l].Clear();
mSumB[l].Clear();
}
RandomizeParameters();
for (int i = 0; i < epochs; i++)
{
for (int j = 0; j < data.TrainImages.ColumnCount; j++)
{
SetInputLayer(data.TrainImages.Column(j));
FeedForward();
SetOutputLayer((int)(data.TrainLabels[j]));
OutputError();
Backprop();
// Perform summation calculations over the training images.
for (int l = lastLayerIndex; l > 0; l--)
{
mSumW[l] += mError[l].ToColumnMatrix() * mNeurons[l - 1].ToRowMatrix();
mSumB[l] += mError[l];
}
// Update the weights and biases after the mini-batch has been processed.
if (j % batchSize == 0 && j > 0)
{
for (int l = 0; l < vSizes.Count; l++)
{
mWeights[l] += mSumW[l].Multiply(-learningRate / batchSize);
mBiases[l] += mSumB[l].Multiply(-learningRate / batchSize);
mSumW[l].Clear();
mSumB[l].Clear();
}
}
}
if (output == true)
{
Console.WriteLine("Epoch {0}: {1} / {2}", i, MnistTest(ref data), data.TestLabels.Count);
}
}
Console.WriteLine("Training Complete.");
}
public int MnistTest(ref MnistData data)
{
int correct = 0;
int total = data.TestLabels.Count;
for (int i = 0; i < total; i++)
{
SetInputLayer(data.TestImages.Column(i));
FeedForward();
if ((mNeurons[lastLayerIndex].MaximumIndex()) == data.TestLabels[i])
{
correct++;
}
}
Console.WriteLine(correct / total);
this.Accuracy = (double)correct / (double)total;
return(correct);
}