public void TrainNN()
{
total = 0;
testIndex = 0;
correct = 0;
nn = new NeuralNetwork(1152, hiddenNeurons, 10);
errors.Clear();
trained = true;
float startTrain = Time.realtimeSinceStartup;
for (int i = 0; i < m_texLoad.trainLength; ++i)
{
// Store errors:
if ((i % 1 == 0 && i > 0) || i == 1)
{
errors.Add(new Point((float)i / m_texLoad.trainLength, nn.cumError / i));
}
int label = m_texLoad.GetLabel(i, 1);
float[] targets = new float[] { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
targets [label] = 1;
Matrix sample = Matrix.fromArray(m_texLoad.GetPixels(i, 1), 28, 28);
Matrix[] features = convLayer.Convolve(sample);
Matrix[] pooledFeatures = convLayer.MaxPool(features);
float [] FC_Array = new float[pooledFeatures.Length * pooledFeatures[0].getCols() * pooledFeatures[0].getRows()];
for (int f = 0; f < 8; ++f)
{
float[] featureMapArray = pooledFeatures [f].toArray();
for (int j = 0; j < featureMapArray.Length; ++j)
{
FC_Array [(featureMapArray.Length * f) + j] = featureMapArray [j];
}
}
// Pass convonlved to FC, performing forward and backprop
nn.train(FC_Array, targets, batchSize);
}
trainTime = Time.realtimeSinceStartup - startTrain;
}
