need help "backpropagation algorithm with more than one hidden layer"

Question

i have implemented backpropagation network for one hidden layer, but now my guide told me to use more than one hidden layer.
can any one tell me:
# what is the method for distribution of neurons in hidden layers(>1)?
# following initialization weight and bias matrices are right for hidden layers=2?
           v=randn(n,x);
            w=randn(x,m);
            wo=randn(1,m);
            vo=randn(1,x);
            v1=randn(x,(p-x)); %2nd hidden layer
            vo1=randn(1,(p-x)); %2nd hidden layer
            vin=v;
            win=w;
            woin=wo;
            voin=vo;
            v1in=v1;  %2nd hidden layer
            vo1in=vo1; %2nd hidden layer
3) whether the following code is correct for 2 hidden layers?
  p-hidden layer neurons
  x=round(p/2)%2 is no. of hidden layers
    %%%First hidden layer
            zin=zeros(trn_dat,x);
            z = zeros(trn_dat,x);
            din = zeros(trn_dat,x);
            d = zeros(trn_dat,x);
            delv = zeros(n,x);
            delvo = zeros(1,x);
            %%%Second hidden layer
            zin1=zeros(x,p-x);
            z1=zeros(x,p-x);
            din1 = zeros(x,p-x);
            d1 = zeros(x,p-x);
            delv1 = zeros(x,p-x);
            delvo1 = zeros(1,p-x);
            %%%Output Layer
            yin = zeros(trn_dat,m);
            y = zeros(trn_dat,m);
            dk = zeros(trn_dat,m);
            delw = zeros(p-x,m);
            delwo = zeros(1,m);
            er=0;
            itr=1;
            max_itr=1000;
            while er==0
                disp('epoch no is');
                disp(itr);
                for T=1:trn_dat  %total no. of training records
                %%feed forward
                    for j=1:x
                        zin(T,j)=0;
                        for i=1:n
                            zin(T,j)=zin(T,j)+(dt(T,i)*v(i,j));
                        end
                    zin(T,j)=zin(T,j)+vo(j);
                    z(T,j)=(2/(1+exp(-zin(T,j))))-1;
                    end
                    for j=(p-x)
                        zin1(j)=0;
                        for i=1:x
                            zin1(j)=zin1(j)+(z(T,i)*v1(i,j));
                        end
                        zin1(j)=zin1(j)+vo1(j);
                        z1(j)=(2/(1+exp(-zin1(j))))-1;
                    end
                    for k=1:m
                        yin(T,k)=0;
                        for j=1:(p-x)
                            yin(T,k)=yin(T,k)+(z1(j)*w(j,k));
                        end
                        yin(T,k)=yin(T,k)+w(k);
                        y(T,k)=(2/(1+exp(-yin(T,k))))-1;
                    end
                    %%back propagation of error
                    for k=1:m
                        dk(T,k)=0;
                        dk(T,k)=((tar(T,k)-y(T,k))*((1/2)*(1+y(T,k))*(1-y(T,k))));
                        for j=1:(p-x)
                            delw(j,k)=(alpha*dk(T,k)*z1(j));
                        end
                        delwo(k)=(alpha*dk(T,k));
                    end
                    for j=1:(p-x)
                        din1(j)=0;
                        for k=1:m
                            din1(j)=din1(j)+((dk(T,k)*w(j,k)));
                        end
                        d1(j)=0;
                        d1(j)=(din1(j)*((1/2)*(1+z1(j))*(1-z1(j))));

                        for i=1:x
                            delv1(i,j)=(alpha*d1(j)*z(T,j));
                        end
                        delvo1(j)=(alpha*d1(j));
                        for i=1:n
                            delv(i,j)=alpha*d(T,j)*dt(T,i);
                        end
                        delvo(j)=alpha*d(T,j);
                    end
                    %%updation of weights and biases
                    for k=1:m
                        for j=1:(p-x)
                            w(j,k)=w(j,k)+delw(j,k);
                        end
                        wo(k)=wo(k)+delwo(k);
                    end
                    for j=1:(p-x)
                        for i=1:x
                            v1(i,j)=v1(i,j)+delv1(i,j);
                        end
                        vo1(j)=vo1(j)+delvo1(j);
                    end
                    for j=1:x
                        for i=1:n
                            v(i,j)=v(i,j)+delv(i,j);
                        end
                        vo(j)=vo(j)+delvo(j);
                    end
                end
                disp('value of y at this iteration:');
                disp(y);
                %    error=0;
                e=tt-y;
                error=sqrt(e.^2);
                if (max(max(error))<0.05)
                    er=1;
                else
                    er=0;
                end
                itr=itr+1;
                if(itr>max_itr)
                    break;
                end
              end

thank you

Punam

need help "backpropagation algorithm with more than one hidden layer"

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

回答（0 个）

另请参阅

类别

标签

Community Treasure Hunt

need help "backpropagation algorithm with more than one hidden layer"

1 个评论 显示 -1更早的评论隐藏 -1更早的评论

回答（0 个）

另请参阅

类别

标签

Community Treasure Hunt

1 个评论
显示 -1更早的评论隐藏 -1更早的评论