static void Main(string[] args)
{
//不使用預設的激發函數組，預設激發函數為LogisticSigmoid
Perceptron.Default_Activation = ActivationFunction.HyperbolicTangent;
Perceptron.Default_DiffActivation = ActivationFunction.Diff_HyperbolicTangent;
//建構一個輸入層數值數量為2 感知器權重亂數範圍為-1~1
//隱藏層節點數量為4 輸出層數量為1的神經網路
PerceptronNetwork network = new PerceptronNetwork(2, -1, 1, new int[] { 4,1 });
LearnData[] data = new LearnData[] {//學習用資料
new LearnData() { Input = new double[] {0,0 },Output = new double[] { 0 }},
new LearnData() { Input = new double[] {1,1 },Output = new double[] { 0 }},
new LearnData() { Input = new double[] {1,0 },Output = new double[] { 1 }},
new LearnData() { Input = new double[] {0,1 },Output = new double[] { 1 }},
};
LearnData.Standardize(data, -1, 1, -1, 1);//將資料轉換為符合Tanh激發函數的資料
//呼叫學習函數，速率0.8，慣性0.6，目標誤差值0.05
network.Train(data,0.8,0.6, 0.05);
//測試
foreach (var item in data) {
Console.WriteLine($"測試({string.Join(",", item.Input)}) => {string.Join(",", network.Compute(item.Input))}");
}
//儲存神經網路結果
StreamWriter writer = new StreamWriter("output.json");
writer.Write(network.ToJObject());//匯出JSON，也可使用Load方法反序列化
writer.Close();
Console.ReadKey();
}
static void Main(string[] args)
{
//不使用預設的激發函數組，預設激發函數為LogisticSigmoid
Perceptron.Default_Activation = ActivationFunction.HyperbolicTangent;
Perceptron.Default_DiffActivation = ActivationFunction.Diff_HyperbolicTangent;
//建構一個輸入層數值數量為2 感知器權重亂數範圍為-1~1
//隱藏層節點數量為4 輸出層數量為1的神經網路
PerceptronNetwork network = new PerceptronNetwork(2, -1, 1, new int[] { 4, 1 });
LearnData[] data = new LearnData[] {//學習用資料
new LearnData()
{
Input = new double[] { 0, 0 }, Output = new double[] { 0 }
},
new LearnData()
{
Input = new double[] { 1, 1 }, Output = new double[] { 0 }
},
new LearnData()
{
Input = new double[] { 1, 0 }, Output = new double[] { 1 }
},
new LearnData()
{
Input = new double[] { 0, 1 }, Output = new double[] { 1 }
},
};
LearnData.Standardize(data, -1, 1, -1, 1);//將資料轉換為符合Tanh激發函數的資料
//呼叫學習函數，速率0.8，慣性0.6，目標誤差值0.05
network.Train(data, 0.8, 0.6, 0.05);
//測試
foreach (var item in data)
{
Console.WriteLine($"測試({string.Join(",", item.Input)}) => {string.Join(",", network.Compute(item.Input))}");
}
//儲存神經網路結果
StreamWriter writer = new StreamWriter("output.json");
writer.Write(network.ToJObject());//匯出JSON，也可使用Load方法反序列化
writer.Close();
Console.ReadKey();
}
/// <summary>
/// 資料集標準化
/// </summary>
/// <param name="Data">原始資料集</param>
/// <param name="InputMin">輸入標準最小值</param>
/// <param name="InputMax">輸入標準最大值</param>
/// <param name="OutputMin">輸出標準最小值</param>
/// <param name="OutputMax">輸出標準最大值</param>
public static void Standardize(LearnData[] Data,int InputMin = 0,int InputMax = 1,int OutputMin = 0,int OutputMax = 1) {
var DataInputMax = Enumerable.Range(0, Data.First().Input.Length).Select(
i => Data.Select(x=>x.Input).Select(x=>x[i]).Max()
).ToArray();
var DataInputMin = Enumerable.Range(0, Data.First().Input.Length).Select(
i => Data.Select(x => x.Input).Select(x => x[i]).Min()
).ToArray();
var DataOutputMax = Enumerable.Range(0, Data.First().Output.Length).Select(
i => Data.Select(x => x.Output).Select(x => x[i]).Max()
).ToArray();
var DataOutputMin = Enumerable.Range(0, Data.First().Output.Length).Select(
i => Data.Select(x => x.Output).Select(x => x[i]).Min()
).ToArray();
for(int i = 0; i < Data.Length; i++) {
Data[i].Input = Data[i].Input.Select((x, i2) =>
ConvertValue((x - DataInputMin[i2]) / (DataInputMax[i2] - DataInputMin[i2]), InputMin, InputMax)
).ToArray();
Data[i].Output = Data[i].Output.Select((x, i2) =>
ConvertValue((x - DataOutputMin[i2]) / (DataOutputMax[i2] - DataOutputMin[i2]), OutputMin, OutputMax)
).ToArray();
}
}
/// <summary>
/// 計算輸入值產生的結果
/// </summary>
/// <param name="Input">學習資料</param>
/// <returns>運算結果</returns>
public double[] Compute(LearnData Input) {
return Compute(Input.Input);
}
/// <summary>
/// 反傳導學習子方法
/// </summary>
/// <param name="Data">學習資料</param>
/// <param name="Rate">學習速率</param>
/// <returns></returns>
private double BackPropagate(LearnPerceptron[][] LearnNodes, LearnData Data, double Rate) {
LearnPerceptron[][] ComputeResult = LearnCompute(Data.Input);
double Error = 0;
#region 感知器節點運算結果誤差
for (int layer = Length - 1; layer > -1; layer--) {
for (int instance = 0; instance < this[layer].Length; instance++) {
if (layer == Length - 1) {//輸出層
LearnNodes[layer][instance].ResultDelta =
(Data.Output[instance] - ComputeResult[layer][instance].Result) *
this[layer][instance].DiffActivation(ComputeResult[layer][instance].Result);
Error += Math.Abs(Data.Output[instance] - ComputeResult[layer][instance].Result);
} else {//隱藏層
//下層偏差值加權總合
double sumDelta = 0;// ComputeResult[layer][instance].Result
for (int nextInstance = 0; nextInstance < Layers[layer + 1].Length; nextInstance++) {
sumDelta += //下一層的節點的誤差加權
LearnNodes[layer + 1][nextInstance].ResultDelta *
this[layer + 1][nextInstance][instance];
}
LearnNodes[layer][instance].ResultDelta = sumDelta *
this[layer][instance].DiffActivation(ComputeResult[layer][instance].Result);
}
}
}
#endregion
#region 計算感知器節點的權重與閥值修正值
for (int layer = Length - 1; layer > -1; layer--) {
for (int instance = 0; instance < this[layer].Length; instance++) {
//修正閥值
LearnNodes[layer][instance].ThresholdDelta += -Rate * LearnNodes[layer][instance].ResultDelta;
for (int weight = 0; weight < this[layer][instance].Length; weight++) {
if (layer == 0) {//輸入層
//本節點偏差與輸入值
for (int upInstance = 0; upInstance < Data.Input.Length; upInstance++) {
LearnNodes[layer][instance].WeightsDelta[upInstance] += Rate * LearnNodes[layer][instance].ResultDelta * Data.Input[upInstance];
}
} else {
//本節點偏差與上層節點給的輸入
for (int upInstance = 0; upInstance < this[layer - 1].Length; upInstance++) {
LearnNodes[layer][instance].WeightsDelta[upInstance] += Rate * LearnNodes[layer][instance].ResultDelta * ComputeResult[layer - 1][upInstance].Result;
}
}
}
}
}
#endregion
return Error;
}
/// <summary>
/// 反傳導學習方法
/// </summary>
/// <param name="Data">學習資料集和</param>
/// <param name="Rate">學習速率</param>
/// <param name="Momentum">慣性動量</param>
/// <param name="Difference">目標誤差值</param>
/// <param name="Iterations">迭代次數限制</param>
public void Train(LearnData[] Data, double Rate, double Momentum = 0, double Difference = 0, int Iterations = -1) {
double Min = double.MaxValue;
for (int i = 0; Iterations == -1 || i < Iterations; i++) {
//學習用暫存資料
LearnPerceptron[][] LearnNodes = GetLearnPerceptrons();
double Error = 0;
foreach (LearnData Item in Data) {
Error += BackPropagate(LearnNodes,Item, Rate);
}
#region 修正
for (int layer = Length - 1; layer > -1; layer--) {
for (int instance = 0; instance < this[layer].Length; instance++) {
//修正閥值
this[layer][instance].Threshold += LearnNodes[layer][instance].ThresholdDelta + LearnNodes[layer][instance].Last_ThresholdDelta * Momentum;// / Math.Sqrt(Data.Length);
for (int weight = 0; weight < this[layer][instance].Length; weight++) {
this[layer][instance][weight] += LearnNodes[layer][instance].WeightsDelta[weight] + LearnNodes[layer][instance].Last_WeightsDelta[weight] * Momentum;// / Math.Sqrt(Data.Length);
}
}
}
#endregion
#region 備份修正值
for(int layer = Length - 1;layer > -1;layer--) {
for(int instance = 0; instance < this[layer].Length; instance++) {
LearnNodes[layer][instance].Last_ThresholdDelta = LearnNodes[layer][instance].ThresholdDelta;
LearnNodes[layer][instance].Last_WeightsDelta = LearnNodes[layer][instance].WeightsDelta;
}
}
#endregion
Min = Math.Min(Min, Error);
Console.WriteLine($"迭代:{i}\t{Error}\t{Min}");
if (Error <= Difference) {
break;
}
}
}
/// <summary>
/// 反傳導學習子方法
/// </summary>
/// <param name="Data">學習資料</param>
/// <param name="Rate">學習速率</param>
/// <returns></returns>
private double BackPropagate(LearnPerceptron[][] LearnNodes, LearnData Data, double Rate)
{
LearnPerceptron[][] ComputeResult = LearnCompute(Data.Input);
double Error = 0;
#region 感知器節點運算結果誤差
for (int layer = Length - 1; layer > -1; layer--)
{
for (int instance = 0; instance < this[layer].Length; instance++)
{
if (layer == Length - 1) //輸出層
{
LearnNodes[layer][instance].ResultDelta =
(Data.Output[instance] - ComputeResult[layer][instance].Result) *
this[layer][instance].DiffActivation(ComputeResult[layer][instance].Result);
Error += Math.Abs(Data.Output[instance] - ComputeResult[layer][instance].Result);
}
else //隱藏層
//下層偏差值加權總合
{
double sumDelta = 0;// ComputeResult[layer][instance].Result
for (int nextInstance = 0; nextInstance < Layers[layer + 1].Length; nextInstance++)
{
sumDelta += //下一層的節點的誤差加權
LearnNodes[layer + 1][nextInstance].ResultDelta *
this[layer + 1][nextInstance][instance];
}
LearnNodes[layer][instance].ResultDelta = sumDelta *
this[layer][instance].DiffActivation(ComputeResult[layer][instance].Result);
}
}
}
#endregion
#region 計算感知器節點的權重與閥值修正值
for (int layer = Length - 1; layer > -1; layer--)
{
for (int instance = 0; instance < this[layer].Length; instance++)
{
//修正閥值
LearnNodes[layer][instance].ThresholdDelta += -Rate * LearnNodes[layer][instance].ResultDelta;
for (int weight = 0; weight < this[layer][instance].Length; weight++)
{
if (layer == 0) //輸入層
//本節點偏差與輸入值
{
for (int upInstance = 0; upInstance < Data.Input.Length; upInstance++)
{
LearnNodes[layer][instance].WeightsDelta[upInstance] += Rate * LearnNodes[layer][instance].ResultDelta * Data.Input[upInstance];
}
}
else
{
//本節點偏差與上層節點給的輸入
for (int upInstance = 0; upInstance < this[layer - 1].Length; upInstance++)
{
LearnNodes[layer][instance].WeightsDelta[upInstance] += Rate * LearnNodes[layer][instance].ResultDelta * ComputeResult[layer - 1][upInstance].Result;
}
}
}
}
}
#endregion
return(Error);
}
/// <summary>
/// 計算輸入值產生的結果
/// </summary>
/// <param name="Input">學習資料</param>
/// <returns>運算結果</returns>
public double[] Compute(LearnData Input)
{
return(Compute(Input.Input));
}