public NeuralLayer(int size, int previousSize, bool activation = false)
{
nodes = new NeuralNode[size];
for (int i = 0; i < size; i++)
{
nodes[i] = new NeuralNode(previousSize, activation);
//Testing
}
}
public void Train(float[] inputData, float[] expectedOutputData)
{
// for (int o = 0; o < 30; o++)
{
timesTrained++;
List <float[]> forwardfeed = new List <float[]>(),
rawForwardFeed = new List <float[]>();
List <float[, ]> newWeights = new List <float[, ]>();
float[] lastOutput = inputData;
for (int i = 0; i < layers.Length; i++)
{
lastOutput = layers[i].Evaluate(lastOutput);
forwardfeed.Add(lastOutput);
}
rawForwardFeed.Add(layers[0].EvaluateRaw(inputData));
for (int i = 1; i < layers.Length; i++)
{
rawForwardFeed.Add(layers[i].EvaluateRaw(forwardfeed[i - 1]));
}
int currentLayer = layers.Length - 1, lastLayer = layers.Length - 1;
//Last layer
float[,] newW = new float[layers[currentLayer].nodes.Length, layers[currentLayer].nodes[0].weight.Length];
for (int i = 0; i < layers[currentLayer].nodes.Length; i++)
{
NeuralNode node = layers[currentLayer].nodes[i];
for (int j = 0; j < node.weight.Length; j++)
{
float newWeight = DeriativeOfCrossEntropy(expectedOutputData[i], forwardfeed[currentLayer][i])
* Neural.TanhDer(rawForwardFeed[currentLayer][i])
* forwardfeed[currentLayer - 1][j];
newW[i, j] = node.weight[j] - Neural.learnRate * newWeight;
}
}
newWeights.Insert(0, newW);
//Rest
currentLayer--;
newW = new float[layers[currentLayer].nodes.Length, layers[currentLayer].nodes[0].weight.Length];
// List<float> errors = new List<float>();
for (int i = 0; i < layers[currentLayer].nodes.Length; i++)
{
NeuralNode node = layers[currentLayer].nodes[i];
for (int j = 0; j < node.weight.Length; j++)
{
float totalError = 0;
for (int k = 0; k < layers[layers.Length - 1].nodes.Length; k++)
{
totalError += DeriativeOfCrossEntropy(expectedOutputData[k], forwardfeed[lastLayer][k])
* Neural.TanhDer(rawForwardFeed[lastLayer][k])
* layers[lastLayer].nodes[k].weight[i];
}
// errors.Add(totalError);
float newWeight = totalError
* Neural.TanhDer(rawForwardFeed[currentLayer][i])
* inputData[j]; //forwardfeed[currentLayer-1][j];
newW[i, j] = node.weight[j] - Neural.learnRate * newWeight;
}
}
newWeights.Insert(0, newW);
/* currentLayer--;
* newW = new float[layers[currentLayer].nodes.Length, layers[currentLayer].nodes[0].weight.Length];
* for (int i = 0; i < layers[currentLayer].nodes.Length; i++)
* {
* NeuralNode node = layers[currentLayer].nodes[i];
* for (int j = 0; j < node.weight.Length; j++)
* {
* float totalError = 0;
* for (int k = 0; k < layers[layers.Length - 1].nodes.Length; k++)
* {
* totalError += errors[k]
* Neural.TanhDer(rawForwardFeed[currentLayer+1][k])
* layers[currentLayer+1].nodes[k].weight[i];
* }
* float newWeight = totalError
* Neural.TanhDer(rawForwardFeed[currentLayer][i])
* inputData[j];
* newW[i, j] = node.weight[j]-Neural.learnRate * newWeight;
* }
* }
* newWeights.Insert(0, newW);*/
for (int i = 0; i < newWeights.Count; i++)
{
newW = newWeights[i];
layers[i].UpdateNodes(newW);
}
}
}
