void Start()
{
observation_inputs = new List <float[, , ]>();
decision_outputs = new List <float[]>();
decision_weights = new List <float>();
evalSolver = Noedify.CreateSolver();
// PlayerPrefs.DeleteAll();
rdb = gameObject.GetComponent <Rigidbody>();
}
void Start()
{
trainingSolver = Noedify.CreateSolver();
trainingSet = new List <Hoppy_Brain.PlayerObservationDecision>();
BuildNetwork();
no_players = 1;
ChangeNoPlayers(false);
lastObstacleSpawn_t = -obstacleSpawnTime;
topScore = 0;
}
// Start is called before the first frame update
void Start()
{
if (simController == null)
{
simController = GameObject.Find("SimController").GetComponent <Hoppy_SimController>();
}
observation_inputs = new List <float[, , ]>();
decision_outputs = new List <float[]>();
decision_weights = new List <float>();
lastMovement_t = -decisionPeriod;
evalSolver = Noedify.CreateSolver();
modifiedDecisionPeriod = decisionPeriod;
}
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();
}
}
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));
}
void Start()
{
trainingSolver = Noedify.CreateSolver();
trainingSet = new List <Hoppy_Brain.PlayerObservationDecision>();
BuildNetwork();
}
Noedify.Net BuildAndImportModel()
{
modelImportComplete = false;
int no_labels = 10;
bool importTensorflowModel = false;
net = new Noedify.Net();
// Attempt to load network saved as a binary file
// This is much faster than importing form a parameters file
bool status = net.LoadModel("Noedify-Model_Digit_Drawing_Test");
if (status == false)
{
print("Binary file not found. Importing Tensorflow parameters.");
importTensorflowModel = true;
}
// If the binary file doesn't exist yet, import the parameters from the Tensorflow file
// This is slower, so we will save the network as a binary file after importing it
if (importTensorflowModel)
{
/* Input layer */
Noedify.Layer inputLayer = new Noedify.Layer(
Noedify.LayerType.Input2D, // layer type
new int[2] {
28, 28
}, // input size
1, // # of channels
"input layer" // layer name
);
net.AddLayer(inputLayer);
// Hidden layer 0
Noedify.Layer hiddenLayer0 = new Noedify.Layer(
Noedify.LayerType.FullyConnected, // layer type
600, // layer size
Noedify.ActivationFunction.Sigmoid, // activation function
"fully connected 1" // layer name
);
net.AddLayer(hiddenLayer0);
// Hidden layer 2
Noedify.Layer hiddenLayer1 = new Noedify.Layer(
Noedify.LayerType.FullyConnected, // layer type
300, // layer size
Noedify.ActivationFunction.Sigmoid, // activation function
"fully connected 2" // layer name
);
net.AddLayer(hiddenLayer1);
// Hidden layer 2
Noedify.Layer hiddenLayer2 = new Noedify.Layer(
Noedify.LayerType.FullyConnected, // layer type
140, // layer size
Noedify.ActivationFunction.Sigmoid, // activation function
"fully connected 3" // layer name
);
net.AddLayer(hiddenLayer2);
/* Output layer */
Noedify.Layer outputLayer = new Noedify.Layer(
Noedify.LayerType.Output, // layer type
no_labels, // layer size
Noedify.ActivationFunction.Sigmoid, // activation function
"output layer" // layer name
);
net.AddLayer(outputLayer);
net.BuildNetwork();
status = NSAI_Manager.ImportNetworkParameters(net, "FC_mnist_600x300x140_parameters");
if (status)
{
print("Successfully loaded model.");
}
else
{
print("Tensorflow model load failed. Have you moved the \"...Assets/Noedify/Resources\" folder to: \"...Assets/Resources\" ?");
print("All model parameter files must be stored in: \"...Assets/Resources/Noedify/ModelParameterFiles\"");
return(null);
}
net.SaveModel("Noedify-Model_Digit_Drawing_Test");
print("Saved binary model file \"Noedify-Model_Digit_Drawing_Test\"");
}
solver = Noedify.CreateSolver();
solver.suppressMessages = true;
modelImportComplete = true;
return(net);
}
