public FeedForwardResult FeedForward(double[] inputs)
{
if (inputs.Length != NumInputs)
{
throw new NeuralNetworkException($"Expected input length to be {NumInputs} but got {inputs.Length}.");
}
var result = new FeedForwardResult();
var inputsWithBias = GetInputsWithBias(inputs);
result.InputWithBias = inputsWithBias;
var hidden = new double[NumHidden + 1];
hidden[NumHidden] = 1; // Bias.
result.HiddenLayers = new [] { hidden };
for (var j = 0; j < NumHidden; j++)
{
var preActivation = 0.0;
var offset = j * (NumInputs + 1);
for (var i = 0; i <= NumInputs; i++)
{
preActivation += HiddenWeights[offset + i] * inputsWithBias[i];
}
hidden[j] = Math.Tanh(preActivation);
}
var sumOfPreOutputs = 0.0;
var preOutputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
{
var preOutput = 0.0;
var offset = k * (NumHidden + 1);
for (var j = 0; j < NumHidden + 1; j++)
{
preOutput += OutputWeights[offset + j] * hidden[j];
}
preOutputs[k] = Math.Exp(preOutput);
sumOfPreOutputs += preOutputs[k];
}
var outputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
{
outputs[k] = preOutputs[k] / sumOfPreOutputs;
}
result.Output = outputs;
return(result);
}
public FeedForwardResult FeedForward(double[] inputs)
{
if (inputs.Length != NumInputs)
throw new NeuralNetworkException($"Expected input length to be {NumInputs} but got {inputs.Length}.");
var result = new FeedForwardResult();
var inputsWithBias = GetInputsWithBias(inputs);
result.InputWithBias = inputsWithBias;
var hidden = new double[NumHidden + 1];
hidden[NumHidden] = 1; // Bias.
result.HiddenLayers = new [] {hidden};
for (var j = 0; j < NumHidden; j++)
{
var preActivation = 0.0;
var offset = j * (NumInputs + 1);
for (var i = 0; i <= NumInputs; i++)
preActivation += HiddenWeights[offset + i] * inputsWithBias[i];
hidden[j] = Math.Tanh(preActivation);
}
var sumOfPreOutputs = 0.0;
var preOutputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
{
var preOutput = 0.0;
var offset = k * (NumHidden + 1);
for (var j = 0; j < NumHidden + 1; j++)
preOutput += OutputWeights[offset + j] * hidden[j];
preOutputs[k] = Math.Exp(preOutput);
sumOfPreOutputs += preOutputs[k];
}
var outputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
outputs[k] = preOutputs[k] / sumOfPreOutputs;
result.Output = outputs;
return result;
}
/// <summary>
/// Activation functions: tanh for hidden nodes, softmax for output nodes.
/// </summary>
public FeedForwardResult FeedForward(double[] inputs)
{
if (inputs.Length != NumInputs)
throw new NeuralNetworkException($"Argument 'inputs' should have width {NumInputs}; was {inputs.Length}.");
// Add bias input node (always = 1).
var inputsWithBias = new double[NumInputs + 1];
Array.Copy(inputs, inputsWithBias, inputs.Length);
inputsWithBias[inputsWithBias.Length - 1] = 1;
// Prepare hidden nodes. Include bias node (always = 1) as the last node in each layer.
var hidden = new double[HiddenLayerSizes.Count][];
for (var i = 0; i < HiddenLayerSizes.Count; i++)
{
var length = HiddenLayerSizes[i] + 1;
hidden[i] = new double[length];
hidden[i][length - 1] = 1;
}
// Calculate first hidden layer.
for (var j = 0; j < HiddenLayerSizes[0]; j++)
{
var preActivation = 0.0;
var offset = j * (NumInputs + 1);
for (var i = 0; i < NumInputs + 1; i++)
preActivation += inputsWithBias[i] * Weights[0][offset + i];
hidden[0][j] = Math.Tanh(preActivation);
}
// Calculate the rest of the hidden layers.
for (var l = 1; l < HiddenLayerSizes.Count; l++)
{
for (var j = 0; j < HiddenLayerSizes[l]; j++)
{
var preActivation = 0.0;
var prevLayerSize = HiddenLayerSizes[l - 1] + 1;
var offset = j * prevLayerSize;
for (var i = 0; i < prevLayerSize; i++)
preActivation += hidden[l-1][i] * Weights[l][offset + i];
hidden[l][j] = Math.Tanh(preActivation);
}
}
// Calculate the output layer.
var lastHiddenLayer = hidden[HiddenLayerSizes.Count - 1];
var lastHiddenLayerSize = lastHiddenLayer.Length;
var outputWeights = Weights[Weights.Length - 1];
var sumOfPreOutputs = 0.0;
var preOutputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
{
var preOutput = 0.0;
var offset = k * lastHiddenLayerSize;
for (var j = 0; j < lastHiddenLayerSize; j++)
preOutput += outputWeights[offset + j] * lastHiddenLayer[j];
preOutputs[k] = Math.Exp(preOutput);
sumOfPreOutputs += preOutputs[k];
}
var outputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
outputs[k] = preOutputs[k] / sumOfPreOutputs;
var result = new FeedForwardResult();
result.InputWithBias = inputsWithBias;
result.HiddenLayers = hidden;
result.Output = outputs;
return result;
}
/// <summary>
/// Activation functions: tanh for hidden nodes, softmax for output nodes.
/// </summary>
public FeedForwardResult FeedForward(double[] inputs)
{
if (inputs.Length != NumInputs)
{
throw new NeuralNetworkException($"Argument 'inputs' should have width {NumInputs}; was {inputs.Length}.");
}
// Add bias input node (always = 1).
var inputsWithBias = new double[NumInputs + 1];
Array.Copy(inputs, inputsWithBias, inputs.Length);
inputsWithBias[inputsWithBias.Length - 1] = 1;
// Prepare hidden nodes. Include bias node (always = 1) as the last node in each layer.
var hidden = new double[HiddenLayerSizes.Count][];
for (var i = 0; i < HiddenLayerSizes.Count; i++)
{
var length = HiddenLayerSizes[i] + 1;
hidden[i] = new double[length];
hidden[i][length - 1] = 1;
}
// Calculate first hidden layer.
for (var j = 0; j < HiddenLayerSizes[0]; j++)
{
var preActivation = 0.0;
var offset = j * (NumInputs + 1);
for (var i = 0; i < NumInputs + 1; i++)
{
preActivation += inputsWithBias[i] * Weights[0][offset + i];
}
hidden[0][j] = Math.Tanh(preActivation);
}
// Calculate the rest of the hidden layers.
for (var l = 1; l < HiddenLayerSizes.Count; l++)
{
for (var j = 0; j < HiddenLayerSizes[l]; j++)
{
var preActivation = 0.0;
var prevLayerSize = HiddenLayerSizes[l - 1] + 1;
var offset = j * prevLayerSize;
for (var i = 0; i < prevLayerSize; i++)
{
preActivation += hidden[l - 1][i] * Weights[l][offset + i];
}
hidden[l][j] = Math.Tanh(preActivation);
}
}
// Calculate the output layer.
var lastHiddenLayer = hidden[HiddenLayerSizes.Count - 1];
var lastHiddenLayerSize = lastHiddenLayer.Length;
var outputWeights = Weights[Weights.Length - 1];
var sumOfPreOutputs = 0.0;
var preOutputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
{
var preOutput = 0.0;
var offset = k * lastHiddenLayerSize;
for (var j = 0; j < lastHiddenLayerSize; j++)
{
preOutput += outputWeights[offset + j] * lastHiddenLayer[j];
}
preOutputs[k] = Math.Exp(preOutput);
sumOfPreOutputs += preOutputs[k];
}
var outputs = new double[NumOutputs];
for (var k = 0; k < NumOutputs; k++)
{
outputs[k] = preOutputs[k] / sumOfPreOutputs;
}
var result = new FeedForwardResult();
result.InputWithBias = inputsWithBias;
result.HiddenLayers = hidden;
result.Output = outputs;
return(result);
}
