void network::SGD(train_couple* mini_batch, int batch_size, double eta,int n)
{
matrix* X=extract_X(mini_batch,batch_size);
matrix* Y=extract_Y(mini_batch,batch_size);
matrix* Gw_sum=prepare_Gw();
matrix* Gb_sum=prepare_Gb();
for (int i=0; i<batch_size; i++){
matrix* Gw=prepare_Gw();
matrix* Gb=prepare_Gb();
back_propagation(X[i],Y[i],Gw,Gb);
for (int l=0; l<layers_count; l++){
Gw_sum[l]=Gw_sum[l]+Gw[l];
Gb_sum[l]=Gb_sum[l]+Gb[l];
}
delete [] Gw;
delete [] Gb;
}
/////Averaging
for (int l=0; l<layers_count; l++){
Gw_sum[l]=Gw_sum[l]*(n/batch_size);
Gb_sum[l]=Gb_sum[l]*(n/batch_size);
}
for (int l=0; l<layers_count; l++){
w[l]=w[l]-eta*Gw_sum[l] ;
b[l]=b[l]-eta*Gb_sum[l] ;
}
delete [] X;
delete [] Y;
delete [] Gw_sum;
delete [] Gb_sum;
}
// betatp => T x p ??
// Xtp => N x p
// XB => N x 1
void comb_XB_tp(int *n, int *r, int *T, int *p, double *Xtp, double *betatp,
int *constant, double *XB)
{
int t, l, n1, r1, T1, p1;
n1 =*n;
r1 =*r;
T1 =*T;
p1 =*p;
double *X1, *beta, *XB1;
X1 = (double *) malloc((size_t)((n1*p1)*sizeof(double)));
beta = (double *) malloc((size_t)((p1)*sizeof(double)));
XB1 = (double *) malloc((size_t)((n1)*sizeof(double)));
for(l=0; l<r1; l++){
for(t=0; t<T1; t++){
extract_X(t, l, n, r, T, p, Xtp, X1); // nrT x p into n x p
extract_beta_t(t, T, p, betatp, beta); // p x T into p x 1
MProd(beta, constant, p, X1, n, XB1); // n x 1
put_together1(l, t, n, r, T, XB, XB1);
// int i;
// for(i=0; i< n1*p1; i++){
// Rprintf(" X1: %4.4f, \n", X1[i]);
// }
// for(i=0; i< p1; i++){
// Rprintf(" beta: %4.4f, \n", beta[i]);
// }
// for(i=0; i< n1; i++){
// Rprintf(" XB: %4.4f, \n", XB1[i]);
// }
}
}
free(X1); free(beta); free(XB1);
return;
}
// Posterior distribution for "theta"
void beta_gp(int *n, int *r, int *T, int *rT, int *p, double *prior_mu,
double *prior_sig, double *Qeta, double *X, double *o, int *constant,
double *betap)
{
int t, l, i, n1, p1, r1, T1, col;
n1 =*n;
p1 =*p;
r1 =*r;
T1 =*T;
col =*constant;
double *del, *chi, *ot1, *X1, *tX1, *out, *tX1QX1, *tX1Qo, *det, *mu, *I;
del = (double *) malloc((size_t)((p1*p1)*sizeof(double)));
chi = (double *) malloc((size_t)((p1*col)*sizeof(double)));
ot1 = (double *) malloc((size_t)((n1*col)*sizeof(double)));
X1 = (double *) malloc((size_t)((n1*p1)*sizeof(double)));
tX1 = (double *) malloc((size_t)((n1*p1)*sizeof(double)));
out = (double *) malloc((size_t)((n1*p1)*sizeof(double)));
tX1QX1 = (double *) malloc((size_t)((p1*p1)*sizeof(double)));
tX1Qo = (double *) malloc((size_t)((p1*col)*sizeof(double)));
det = (double *) malloc((size_t)((col)*sizeof(double)));
mu = (double *) malloc((size_t)((p1*col)*sizeof(double)));
I = (double *) malloc((size_t)((p1*p1)*sizeof(double)));
for(i=0; i<p1*p1; i++){
del[i] = 0.0;
}
for(i=0; i<p1; i++){
chi[i] = 0.0;
}
for(l=0; l<r1; l++){
for(t=0; t<T1; t++){
extract_X(t, l, n, r, T, p, X, X1); // n x p
MTranspose(X1, p, n, tX1); // p x n
MProd(X1, p, n, Qeta, n, out); // n x p
MProd(out, p, n, tX1, p, tX1QX1); // pxp
MAdd(del, p, p, tX1QX1, del); // pxp
extract_alt2(l, t, n, rT, T, o, ot1); // n x 1
MProd(ot1, constant, n, Qeta, n, out); // n x 1
MProd(out, constant, n, tX1, p, tX1Qo); // p x 1
MAdd(chi, p, constant, tX1Qo, chi); // p x 1
}
}
IdentityM(p, I);
for(i=0; i<p1*p1; i++){
del[i] = del[i] + I[i]*(1.0/prior_sig[0]);
}
free(I);
for(i=0; i<p1; i++){
chi[i] = chi[i] + prior_mu[0]/prior_sig[0];
}
MInv(del, del, p, det);
MProd(chi, constant, p, del, p, mu); // p x 1
mvrnormal(constant, mu, del, p, betap);
free(del); free(chi); free(ot1); free(X1); free(tX1);
free(out); free(tX1QX1); free(tX1Qo); free(det); free(mu);
return;
}
