public void gradientWeights()
{
const int inSize = 3;
const int outSize = 2;
const int countExamples = 1;
var input = MatrixD.Build.DenseOfArray(
new double[countExamples, inSize] {
{ 1, 2, 3 }
});
var nextGradient = MatrixD.Build.DenseOfArray(
new double[countExamples, outSize] {
{ 4, 5 }
});
var weights = MatrixD.Build.DenseOfArray(
new double[inSize, outSize] {
{ -1, 2 },
{ -3, 2 },
{ -1, 3 },
});
var expectedGradient = MatrixD.Build.DenseOfArray(
new double[inSize, outSize] {
{ 4, 5 },
{ 8, 10 },
{ 12, 15 },
});
var layer = new WeightLayer(inSize, outSize);
var gradient = layer.gradientWeights(input, nextGradient);
Assert.IsTrue(gradient.EEquals(expectedGradient));
}
public void forward()
{
const int inSize = 3;
const int outSize = 2;
const int countExamples = 1;
var input = MatrixD.Build.DenseOfArray(
new double[countExamples, inSize] {
{ 1, 2, 3 }
});
var weights = MatrixD.Build.DenseOfArray(
new double[inSize, outSize] {
{ -1, 2 },
{ -3, 2 },
{ -1, 3 },
});
var expectedOutput = MatrixD.Build.DenseOfArray(
new double[countExamples, outSize] {
{ -10, 15 }
});
var layer = new WeightLayer(weights);
var output = layer.forward(input);
Assert.IsTrue(output.EEquals(expectedOutput));
}
/// <summary> Creates, loads, and returns a <see cref="NeuralNetwork"/> using the given <paramref name="networkLoader"/>. </summary>
/// <param name="networkLoader"> The <see cref="INetworkLoader"/> used to load the network. </param>
/// <returns> The loaded <see cref="NeuralNetwork"/>. </returns>
public static NeuralNetwork Load(INetworkLoader networkLoader)
{
// Get the number of neuron layers in the file.
int neuronLayerCount = networkLoader.GetNetworkNeuronLayerCount();
// Create an empty neural network.
NeuralNetwork neuralNetwork = new NeuralNetwork(networkLoader.GetNetworkLearningRate(), neuronLayerCount);
// Create the neuron layers for the neural network.
for (uint i = 0; i < neuronLayerCount; i++)
{
neuralNetwork.neuronLayers[i] = NeuronLayer.Load(neuralNetwork, networkLoader, i);
}
// Create the weight layers for the neural network.
for (uint i = 0; i < neuronLayerCount - 1; i++)
{
neuralNetwork.weightLayers[i] = WeightLayer.Load(neuralNetwork, networkLoader, i, neuralNetwork.neuronLayers[i], neuralNetwork.neuronLayers[i + 1]);
}
// Set the highest neuron count.
neuralNetwork.HighestNeuronCount = calculateHighestNeuronCount(neuralNetwork);
// Return the neural network.
return(neuralNetwork);
}
/// <summary> Creates and initialises each <see cref="NeuronLayer"/> in the <see cref="neuronLayers"/>. </summary>
/// <param name="random"> The <see cref="Random"/> object used to randomise the weights and biases of the network. </param>
/// <param name="layerSizes"> The desired sizes of each <see cref="NeuronLayer"/>. </param>
private void initialiseLayers(Random random, int[] layerSizes)
{
// Create the neuron layers.
for (uint i = 0; i < layerSizes.Length; i++)
{
neuronLayers[i] = new NeuronLayer(this, i, layerSizes[i], random);
}
// Create the weight layers. Note that the input and output layers only have one weight layer.
for (uint i = 0; i < layerSizes.Length - 1; i++)
{
// Create the weight layer.
weightLayers[i] = new WeightLayer(this, i, neuronLayers[i], neuronLayers[i + 1]);
// Link the layers.
weightLayers[i].LinkAllNeurons(random);
}
}
public void backwardLearn()
{
const double learnRate = 0.5;
const int inSize = 3;
const int outSize = 2;
const int countExamples = 1;
var input = MatrixD.Build.DenseOfArray(
new double[countExamples, inSize] {
{ 1, 2, 3 }
});
var nextGradient = MatrixD.Build.DenseOfArray(
new double[countExamples, outSize] {
{ 4, 5 }
});
var weights = MatrixD.Build.DenseOfArray(
new double[inSize, outSize] {
{ -1, 2 },
{ -3, 2 },
{ -1, 3 },
});
var expectedWeights = MatrixD.Build.DenseOfArray(
new double[inSize, outSize] {
{ -3, -0.5 },
{ -7, -3 },
{ -7, -4.5 },
});
var layer = new WeightLayer(weights);
layer.backwardLearn(input, nextGradient, learnRate);
Assert.IsTrue(expectedWeights.EEquals(layer.Weights));
}
