static PetscErrorCode bfgs_apply(PC pc, Vec xin, Vec xout)
{
TaoLMVMMat *M ;
TaoVecPetsc Xin(xin);
TaoVecPetsc Xout(xout);
TaoTruth info2;
int info;
PetscFunctionBegin;
PetscTruth VerbosePrint = PETSC_FALSE;
PetscOptionsGetTruth(PETSC_NULL,"-verboseapp",&VerbosePrint,PETSC_NULL);
info = PCShellGetContext(pc,(void**)&M); CHKERRQ(info);
PetscScalar solnNorm,solnDot;
info = VecNorm(xin,NORM_2,&solnNorm); CHKERRQ(info)
info=PetscPrintf(PETSC_COMM_WORLD,"bfgs_apply: ||Xin||_2 = %22.15e\n",solnNorm);
if(VerbosePrint) VecView(xin,0);
info = M->Solve(&Xin, &Xout, &info2); CHKERRQ(info);
info = VecNorm(xout,NORM_2,&solnNorm); CHKERRQ(info)
info = VecDot(xin,xout,&solnDot); CHKERRQ(info)
info=PetscPrintf(PETSC_COMM_WORLD,"bfgs_apply: ||Xout||_2 = %22.15e, Xin^T Xout= %22.15e\n",solnNorm,solnDot);
if(VerbosePrint) VecView(xout,0);
PetscFunctionReturn(0);
}
static double compute_Xin()
{
realvar *tmp;
/* printf("Computing: Xin[]\n"); */
tmp = &(Xin());
if (!tmp->computed) {
printf("Input Xin =" );
gets(__s_);
tmp->value = atof(__s_);
tmp->computed = 1;
}
return tmp->value;
}
void apply_adjoint(XType& X) const {
XType Xin(X);
base::Gemm(elem::ADJOINT, elem::NORMAL, 1.0, N, Xin, X);
}
void apply(XType& X) const {
XType Xin(X);
base::Gemm(elem::NORMAL, elem::NORMAL, 1.0, N, Xin, X);
}
//Algoritmo usado: backpropagation
void RNA::treina(const char* fn, const char* fout, bool isBatch, bool isNature)
{
FILE* f=fopen(fn,"r");
FILE* flog=fopen("treino.log","w+");
if(!f){
fprintf(stderr,"Falha ao abrir %s!\n",fn);
exit(3);
}
int N;
fscanf(f,"%d",&N);
vector<vector<double> > Xin(N);
vector<vector<double> > d(N);//Saida desejada
printf("Carregando banco de treinamento...\n");
for(int i=0;i<N;i++){
Xin[i].resize(nIn); d[i].resize(nOut);
for(int j=0;j<nIn;j++) fscanf(f,"%lf",&Xin[i][j]);
for(int j=0;j<nOut;j++) fscanf(f,"%lf",&d[i][j]);
}
int NI;
printf("Executando treinamento...\n");
fscanf(f,"%d",&NI);
fclose(f);
vector<double> e(nNeu+1);
vector<double> delta(nNeu+1);
vector<vector<double> > dW;
vector<double> err_ant(nOut);
vector<vector<double> > wt1, wt;
if(isBatch||isNature){
dW.resize(W.size());
wt1.resize(W.size());
wt.resize(W.size());
for(int i=0;i<W.size();i++){
dW[i].resize(W[i].size());
wt1[i].resize(W[i].size());
wt[i].resize(W[i].size());
}
//for(int i=0;i<W.size();i++)
//for(int j=0;j<W[i].size();j++)
//W[i][j] += rand(-1, 1);
} else {
wt1.resize(W.size());
wt.resize(W.size());
for(int i=0;i<W.size();i++){
wt1[i].resize(W[i].size());
wt[i].resize(W[i].size());
for (int j=0;j<W[i].size();j++)
wt[i][j]=wt1[i][j]=0;
}
}
bool isf = true;
double TA0 = TA;
for(int kp=0;kp<NI;kp++){
printf("passo %d\n",kp);
double err_acc = 0;
double alpha = rand(0, 1) * 0.01 * TA0;
TA = rand(0.8, 1.2) * TA0;
if (isBatch){
for(int i=0;i<W.size();i++)
for(int j=0;j<W[i].size();j++)
dW[i][j]=0;
}
if (isNature){
double e0, e1;
e0 = 0;
for(int k=0;k<N;k++){
g(Xin[k]);
gd(Xin[k]);
//Atualizando neuronio na ultima camada
for(int j=0;j<nOut;j++) {
e[cam[NC-1][j]]=d[k][j]-G[cam[NC-1][j]];
e0 += e[cam[NC-1][j]] * e[cam[NC-1][j]];
err_acc+= e[cam[NC-1][j]] * e[cam[NC-1][j]];
}
}
e0=e0/(2);
for(int i=1;i<W.size();i++)
for(int j=1;j<W[i].size();j++){
dW[i][j]=TA*rand(-10,10);
W[i][j]+=dW[i][j];
}
e1 = 0;
for(int k=0;k<N;k++){
g(Xin[k]);
gd(Xin[k]);
//Atualizando neuronio na ultima camada
for(int j=0;j<nOut;j++) {
e[cam[NC-1][j]]=d[k][j]-G[cam[NC-1][j]];
e1 += e[cam[NC-1][j]] * e[cam[NC-1][j]];
}
}
e1=e1/(2);
if(e1>e0){
for(int i=1;i<W.size();i++)
for(int j=1;j<W[i].size();j++){
W[i][j]-=dW[i][j];
}
}
}else{
//Apresentando a rede a todas amostras
for(int k=0;k<N;k++){
g(Xin[k]);
gd(Xin[k]);
//Atualizando neuronio na ultima camada
for(int j=0;j<nOut;j++) {
e[cam[NC-1][j]]=d[k][j]-G[cam[NC-1][j]];
err_acc += e[cam[NC-1][j]] * e[cam[NC-1][j]];
}
for(int j=0;j<nOut;j++){
int yj, i;
yj=cam[NC-1][j];
delta[yj]=e[yj]*Gd[yj];//Gd[yj];
for(int m=0;m<rede[yj].conn.size();m++){
i=rede[yj].conn[m];
double dw = TA*delta[yj]*G[i];
double dwt = 0;
if (isf) {
alpha = 0;
wt[i][yj] = W[i][yj];
} else {
wt1[i][yj] = wt[i][yj];
wt[i][yj] = W[i][yj];
dwt = wt[i][yj] - wt1[i][yj];
}
if(isBatch)dW[i][yj]+= dw;
else W[i][yj]+= dw + alpha * dwt;
}
}
//Atualizando neuronios nas camadas ocultas
for(int l=NC-2;l>-1;l--){
for(int j=0;j<cam[l].size();j++){
int yj, k;
double w;
yj=cam[l][j];
delta[yj]=0;
for(int m=0;m<rede[yj].dest.size();m++){
k=rede[yj].dest[m];
w=W[yj][k];
delta[yj]+=delta[k]*w;
}
delta[yj]*=Gd[yj];//Gd[yj]
for(int m=0;m<rede[yj].conn.size();m++){
int i=rede[yj].conn[m];
double dw = TA*delta[yj]*G[i];
double dwt = 0;
if (isf) {
alpha = 0;
wt[i][yj] = W[i][yj];
} else {
wt1[i][yj] = wt[i][yj];
wt[i][yj] = W[i][yj];
dwt = wt[i][yj] - wt1[i][yj];
}
if(isBatch)dW[i][yj]+= dw;
else W[i][yj]+= dw + alpha * dwt;
}
}
}
if(isBatch){
for(int i=0;i<W.size();i++)
for(int j=0;j<W[i].size();j++)
W[i][j]+=dW[i][j];
}
}
}
fprintf(flog, "%d %e\n", kp,err_acc*0.5);
printf("err=%e\n", err_acc*0.5);
isf = false;
}
printf("Treino efetuado!\n");
fclose(flog);
//Salva a rede
save(fout);
}
