/// <summary>
/// 对特征向量进行分类
/// </summary>
/// <param name="weightMatrixList">权值矩阵列表，对样本进行线性变换</param>
/// <param name="vector">特征列向量</param>
/// <returns>返回分类结果，如果返回-1则表示分类失败！</returns>
public static int Classify(List <decimal[, ]> weightMatrixList, decimal[,] vector)
{
foreach (decimal[,] weightMatrix in weightMatrixList)
{
vector = DefaultActivateFunc.Active(MatrixHelper.MatrixMultiMatrix(weightMatrix, vector));
}
int type = -1;
int count = 0;
for (int i = 0, len = vector.GetLength(0); i < len; i++)
{
if (1M - vector[i, 0] <= .1M)
{
type = i;
count++;
}
}
return(count > 1?-1:type);
}
/// <summary>
/// 训练方法
/// </summary>
/// <param name="weightMatrixList">权值矩阵列表，对样本进行线性变换</param>
/// <param name="matrix">输入矩阵，每一列就是一个样本的特征向量</param>
/// <param name="desiredMatrix">期望矩阵</param>
/// <returns>返回训练后的权值矩阵</returns>
public static List <decimal[, ]> Train(List <decimal[, ]> weightMatrixList, decimal[,] matrix, decimal[,] desiredMatrix)
{
decimal learnRate = new decimal(.11);
List <decimal[, ]> prevInput = new List <decimal[, ]>();
List <decimal[, ]> prevWeight = new List <decimal[, ]>();
List <decimal[, ]> prevOutput = new List <decimal[, ]>();
List <decimal[, ]> prevActive = new List <decimal[, ]>();
decimal[,] output;
foreach (decimal[,] weightMatrix in weightMatrixList)
{
prevWeight.Insert(0, weightMatrix);
prevInput.Insert(0, matrix);
matrix = DefaultActivateFunc.Active(output = weightMatrix.MatrixMultiMatrix(matrix));
prevOutput.Insert(0, output);
prevActive.Insert(0, matrix);
}
decimal[,] backPro = desiredMatrix.MatrixReduceMatrix(prevActive[0]).MatrixVec().VecDiag().MatrixMultiMatrix(
DefaultActivateFunc.Dactive(prevOutput[0]).MatrixVec().VecDiag().MatrixMultiScalar(-1));
List <decimal[, ]> partialWeight = new List <decimal[, ]>();
partialWeight.Insert(0, backPro.MatrixMultiMatrix(prevInput[0].Turn().MatrixKronecker(MatrixHelper.CreateIdentityMatrix(prevWeight[0].GetLength(0)))).Turn());
for (int i = 1, len = weightMatrixList.Count; i < len; i++)
{
backPro = backPro.MatrixMultiMatrix(MatrixHelper.CreateIdentityMatrix(prevActive[i].GetLength(1)).MatrixKronecker(weightMatrixList[i - 1]))
.MatrixMultiMatrix(DefaultActivateFunc.Dactive(prevActive[i]).MatrixVec().VecDiag());
partialWeight.Insert(0, backPro.MatrixMultiMatrix(prevInput[i].Turn().MatrixKronecker(MatrixHelper.CreateIdentityMatrix(weightMatrixList[i].GetLength(0)))));
}
for (int i = 0, len = weightMatrixList.Count; i < len; i++)
{
weightMatrixList[i] = weightMatrixList[i].MatrixAndMatrix(partialWeight[i].MatrixMultiScalar(learnRate));
}
return(weightMatrixList);
}
/// <summary>
/// 克罗内克积
/// </summary>
/// <param name="matrix"></param>
/// <returns></returns>
public Matrix Kronecker(Matrix matrix)
{
return(new Matrix(MatrixHelper.MatrixKronecker(InnerMatrix, matrix.InnerMatrix)));
}
/// <summary>
/// 创建单位矩阵
/// </summary>
/// <param name="width"></param>
public Matrix(int width)
{
InnerMatrix = MatrixHelper.CreateIdentityMatrix(width);
}
/// <summary>
/// 逐元素乘法
/// </summary>
/// <param name="matrix"></param>
/// <returns></returns>
public Matrix Dot(Matrix matrix)
{
return(new Matrix(MatrixHelper.MatrixDotMatrix(InnerMatrix, matrix.InnerMatrix)));
}
