IEnumerator PlotCostWhenComplete(Noedify_Solver solver, float[] cost)
{
while (solver.trainingInProgress)
{
yield return(null);
}
debugger.PlotCost(cost, new float[2] {
1.0f / no_epochs * 2.5f, 5
}, costPlotOrigin);
for (int n = 0; n < predictionTester.sampleImagePlanes.Length; n++)
{
float[,,] testInputImage = new float[1, 1, 1];
Noedify_Utils.ImportImageData(ref testInputImage, predictionTester.sampleImageRandomSet[n], true);
solver.Evaluate(net, testInputImage, Noedify_Solver.SolverMethod.MainThread);
int prediction = Noedify_Utils.ConvertOneHotToInt(solver.prediction);
predictionTester.CNN_predictionText[n].text = prediction.ToString();
}
}
// Start is called before the first frame update
void Start()
{
sampleImageRandomSet = new List <Texture2D>();
int[] sampleRange = new int[sampleImageSet.Length];
for (int i = 0; i < sampleRange.Length; i++)
{
sampleRange[i] = i;
}
Noedify_Utils.Shuffle(sampleRange);
for (int i = 0; i < sampleImagePlanes.Length; i++)
{
sampleImageRandomSet.Add(sampleImageSet[sampleRange[i]]);
}
for (int i = 0; i < sampleImagePlanes.Length; i++)
{
sampleImagePlanes[i].GetComponent <MeshRenderer>().material.SetTexture("_MainTex", sampleImageRandomSet[i]);
}
}
public void TrainModel()
{
List <float[, , ]> trainingData = new List <float[, , ]>();
List <float[]> outputData = new List <float[]>();
List <Texture2D[]> MNIST_images = new List <Texture2D[]>();
MNIST_images.Add(MNIST_images0);
MNIST_images.Add(MNIST_images1);
MNIST_images.Add(MNIST_images2);
MNIST_images.Add(MNIST_images3);
MNIST_images.Add(MNIST_images4);
MNIST_images.Add(MNIST_images5);
MNIST_images.Add(MNIST_images6);
MNIST_images.Add(MNIST_images7);
MNIST_images.Add(MNIST_images8);
MNIST_images.Add(MNIST_images9);
Noedify_Utils.ImportImageData(ref trainingData, ref outputData, MNIST_images, true);
debugger.net = net;
Noedify_Solver.SolverMethod solverMethod = Noedify_Solver.SolverMethod.MainThread;
if (solverMethodToggle != null)
{
if (solverMethodToggle.isOn)
{
solverMethod = Noedify_Solver.SolverMethod.Background;
}
}
if (solver == null)
{
solver = Noedify.CreateSolver();
}
solver.debug = new Noedify_Solver.DebugReport();
sw.Start();
//solver.costThreshold = 0.01f; // Add a cost threshold to prematurely end training when a suitably low error is achieved
//solver.suppressMessages = true; // suppress training messages from appearing in editor the console
solver.TrainNetwork(net, trainingData, outputData, no_epochs, batch_size, trainingRate, costFunction, solverMethod, null, 8);
float[] cost = solver.cost_report;
StartCoroutine(PlotCostWhenComplete(solver, cost));
}
public IEnumerator TrainNetwork()
{
if (trainingSet != null)
{
if (trainingSet.Count > 0)
{
while (trainingSolver.trainingInProgress)
{
yield return(null);
}
List <float[, , ]> observation_inputs = new List <float[, , ]>();
List <float[]> decision_outputs = new List <float[]>();
List <float> trainingSetWeights = new List <float>();
for (int n = 0; n < trainingSet.Count; n++)
{
observation_inputs.Add(Noedify_Utils.AddTwoSingularDims(trainingSet[n].observation));
decision_outputs.Add(trainingSet[n].decision);
trainingSetWeights.Add(trainingSet[n].weight);
}
trainingSolver.TrainNetwork(net, observation_inputs, decision_outputs, trainingEpochs, trainingBatchSize, trainingRate, Noedify_Solver.CostFunction.MeanSquare, Noedify_Solver.SolverMethod.MainThread, trainingSetWeights, N_threads);
}
}
}
// Evaluate netowrk to generate decision
float[] AIDecision(float[] observation)
{
evalSolver.Evaluate(simController.net, Noedify_Utils.AddTwoSingularDims(observation), Noedify_Solver.SolverMethod.MainThread);
return(evalSolver.prediction);
}
