private IEnumerator TrainingProcess()
{
var neuralNetworkConfig = new NetworkDescription()
{
Layers = new List <LayerParamaters>()
{
new LayerParamaters()
{
LayerType = ELayerType.Input,
NeuronCount = InputAttributeCount,
NeuronType = ENeruonType.Input
},
new LayerParamaters()
{
LayerType = ELayerType.Hidden,
NeuronCount = 64,
NeuronType = ENeruonType.Sigmoid
},
new LayerParamaters()
{
LayerType = ELayerType.Hidden,
NeuronCount = 16,
NeuronType = ENeruonType.Sigmoid
},
new LayerParamaters()
{
LayerType = ELayerType.Output,
NeuronCount = 1,
NeuronType = ENeruonType.Sigmoid
},
}
};
var neuralNetwork = new NeuralNetwork(neuralNetworkConfig);
neuralNetwork.InitNetworkWithRandomValues(-0.05f, 0.05f);
var networkEvaluator = new NetworkEvaluator(neuralNetwork);
var tempResult = new NativeArray2D <float>(1, 1);
//Test inital accuracy
TestInitialAccuracy(networkEvaluator, tempResult.Slice());
//Learn over a number of iterations
for (int epoch = 0; epoch < Epochs; epoch++)
{
//For each epoch, perform training
for (int tc = 0; tc < _dataset.TrainingSetSize; tc++)
{
_dataset.GetTrainingCase(tc, out var trainingInput, out var trainingResult);
//Evolve network
var jobHandle = networkEvaluator.GradientDescentBackpropigate(trainingInput, trainingResult, 1, out _);
jobHandle.Complete();
}
float totalError = 0.0f;
int totalCorrect = 0;
for (int i = 0; i < _dataset.TestingSetSize; i++)
{
_dataset.GetTestCase(i, out var trainingInput, out var trainingResult);
networkEvaluator.Evaluate(trainingInput, tempResult.Slice(), 1).Complete();
var error = Math.Abs(tempResult[0, 0] - trainingResult[0, 0]);
bool wasCorrect = error < 0.5f;
totalCorrect += wasCorrect ? 1 : 0;
totalError += error;
}
float averageError = totalError / (float)_dataset.TestingSetSize;
float accuracy = (float)totalCorrect / (float)_dataset.TestingSetSize;
//Forward test
Debug.Log($"Epoch {epoch}: Accuracy:{accuracy:P2} Average Error:{averageError:F4}");
yield return(null);
}
tempResult.BackingStore.Dispose();
}