void backwardTrain(CostLayer costLayer, double learnRate)
{
var gradients = costLayer.backward(currentAnnOutput);
for (int i = layers.Length - 1; i > 0; i--)
{
layers[i].backwardLearn(layerInputs[i], gradients, learnRate);
gradients = layers[i].backward(layerInputs[i], gradients);
}
layers[0].backwardLearn(layerInputs[0], gradients, learnRate);
}
void forward(MatrixD inputs, CostLayer costLayer)
{
layerInputs[0] = inputs;
for (int i = 1; i < this.layers.Length; i++)
{
layerInputs[i] = this.layers[i - 1].forward(layerInputs[i - 1]);
}
this.currentAnnOutput = this.layers.Last().forward(layerInputs.Last());
if (preCost == null)
{
preCost = costLayer.forward(this.currentAnnOutput);
}
else
{
postCost = costLayer.forward(this.currentAnnOutput);
}
//var m = annOutput.ToRowArrays();
}
public void next(double learnRate, int countEntries)
{
preCost = postCost = null;
var randomEntriesIdx = GlobalRandom.NextIntArr(countEntries, 0, annInputs.RowCount - 1);
this.currentInput = annInputs.lines(randomEntriesIdx);
this.currentExpectedOutput = annExpectedOutputs.lines(randomEntriesIdx);
var costLayer = new CostLayer(currentExpectedOutput);
forward(currentInput, costLayer);
backwardTrain(costLayer, learnRate / countEntries);
/* * /
* var mO = annOutput.ToRowArrays();
* var mEO = currentExpectedOutput.ToRowArrays();
*
* var mOidx = annOutput.maxIdxEachRow();
* var mEOidx = currentExpectedOutput.maxIdxEachRow();
* //*/
}
