void BuildNetwork()
{
net = new Noedify.Net();
/* Input layer */
Noedify.Layer inputLayer = new Noedify.Layer(
Noedify.LayerType.Input, // layer type
2, // input size
"antenas layer" // layer name
);
net.AddLayer(inputLayer);
// Hidden layer 1
Noedify.Layer hiddenLayer0 = new Noedify.Layer(
Noedify.LayerType.FullyConnected, // layer type
150, // layer size
Noedify.ActivationFunction.Sigmoid, // activation function
"fully connected 1" // layer name
);
net.AddLayer(hiddenLayer0);
/* Output layer */
Noedify.Layer outputLayer = new Noedify.Layer(
Noedify.LayerType.Output, // layer type
2, // layer size
Noedify.ActivationFunction.Sigmoid, // activation function
"output layer" // layer name
);
net.AddLayer(outputLayer);
net.BuildNetwork();
}
void BuildModel()
{
net = new Noedify.Net();
/* 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 1
Noedify.Layer hiddenLayer0 = new Noedify.Layer(
Noedify.LayerType.Convolutional2D, // layer type
inputLayer, // input to layer
new int[2] {
4, 4
}, // filter size
new int[2] {
2, 2
}, // stride
8, // # of filters
new int[2],
Noedify.ActivationFunction.ReLU, // activation function
"convolutional 1" // layer name
);
net.AddLayer(hiddenLayer0);
// Hidden layer 2
Noedify.Layer hiddenLayer1 = new Noedify.Layer(
Noedify.LayerType.FullyConnected, // layer type
100,
Noedify.ActivationFunction.Sigmoid,
"fully connected 1" // layer name
);
net.AddLayer(hiddenLayer1);
// 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();
}
void TestMNISTDrawing(Noedify.Net net)
{
float[] modelInputs = sampleDrawing.SampleDrawing(new int[2] {
28, 28
});
float[,,] modelInputsFormatted = new float[1, 1, 28 * 28];
for (int i = 0; i < 28 * 28; i++)
{
modelInputsFormatted[0, 0, i] = modelInputs[i];
}
System.Diagnostics.Stopwatch sw = new System.Diagnostics.Stopwatch();
sw.Start();
solver.Evaluate(net, modelInputsFormatted, Noedify_Solver.SolverMethod.MainThread);
sw.Stop();
StartCoroutine(UpdatePlotWhenComplete());
}
void BuildModel()
{
net = new Noedify.Net();
/* 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 1
Noedify.Layer hiddenLayer0 = new Noedify.Layer(
Noedify.LayerType.FullyConnected, // layer type
200,
Noedify.ActivationFunction.Sigmoid,
"fully connected 1" // layer name
);
net.AddLayer(hiddenLayer0);
// Output layer
Noedify.Layer outputLayer = new Noedify.Layer(
Noedify.LayerType.Output, // layer type
no_labels, // layer size
Noedify.ActivationFunction.SoftMax, // activation function
"output layer" // layer name
);
net.AddLayer(outputLayer);
net.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);
}
// Start is called before the first frame update
void Start()
{
mnist_FC_net = BuildAndImportModel();
}
