public MLUnit3(int NA, int NB, MLUnit3 preM, Func <double, double> ActFunc, Func <double, double> BackFunc,
double gamma)
{
ID = ID0++;
if (preM != null)
{
preM.nxtMLU = this; this.preMLU = preM;
}
this.ActFunc = ActFunc; this.BackFunc = BackFunc;
this.NA = NA; this.NB = NB;
U_lst = new DenseVector(NA);
Z_lst = new DenseVector(NA);
D_lst = new DenseVector(NA);
this.gamma = gamma;
if (NB > 0)
{
DropoutLst = new bool[NA];
W_lst = new DenseMatrix(NB, NA);
dW_lst = new DenseMatrix(NB, NA);
pdW_lst = new DenseMatrix(NB, NA);
Learning_Init();
}
WriteLine($"MLUnit3 ID:{ID} NA:{NA} NB:{NB} DropOutEnable:{DropOutEnable}");
}
private void RecursiveCalcu_Propagation(MLUnit3 P, ULeData X, bool DoB, bool DispB)
{
//===== Forward calculation =====
//$$ ApplyDropout_____(P,DoB:P.DropOutEnable);
P.Z_lst[0] = 1.0;
var Q = P.nxtMLU;
Q.U_lst = P.W_lst * P.Z_lst;
Q.U_lst.Apply(Q.Z_lst, x => P.ActFunc(x)); //Activation Function
//===== Backward calculation =====
if (Q.nxtMLU != null) // Intermediate layer
{
RecursiveCalcu_Propagation(Q, X, DoB, DispB); // [Next layer]
}
else // Final layer
//DenseVector E = new DenseVector(Nout,0.0); E.Clear();
//DenseVector E = DenseVector.Create(Nout,0.0); E[X.ans]=1.0;
{
DenseVector E = DenseVector.Create(Nout, p => (p == X.ans)?1.0:0.0);
Q.D_lst = E - Q.Z_lst;
}
DenseVector SU = sigmoidFDash(P.U_lst);
DenseMatrix Wt = (DenseMatrix)P.W_lst.Transpose();
DenseVector WD = Wt * Q.D_lst;
P.D_lst = eProduct(SU, WD); //Hadamard product //②
//$$ ApplyDropout_____(P,DoB:true);
P.dW_lst = vvProduct(Q.D_lst, P.Z_lst);
return;
}
private void ApplyDropout_____(MLUnit3 MLU, bool DoB)
{
if (!dropoutB || !MLU.DropOutEnable)
{
return;
}
double gamma = MLU.gamma;
var Z = MLU.Z_lst;
if (DoB)
{
var B = MLU.DropoutLst;
for (int k = 1; k < MLU.NA; k++)
{
if (B[k])
{
Z[k] = 0.0;
}
}
}
else
{
Z *= gamma;
}
}
private void SetArray_Weight(int[] NSizeLst, double gammaC = 9999.0)
{
this.NSizeLst = NSizeLst;
MLUnit3 preM = null, Q;
MLUnit3.ID0 = 0;
MLUs = new List <MLUnit3>();
for (int n = 0; n < NLayer; n++)
{
int NA = NSizeLst[n];
int NB = (n < NLayerM1)? NSizeLst[n + 1]: 0;
double gamma = gammaC;
if (n == NLayerM1)
{
gamma = 1.0;
}
MLUs.Add(Q = new MLUnit3(NA, NB, preM, sigmoidF, sigmoidFDash, gamma));
preM = Q;
}
}