//--------------------------------------------------------------------------
double Performance(const DVector &ytest_truth, const DMatrix &Ytest_predict,
int NumClasses) {
long n = ytest_truth.GetN();
if (n != Ytest_predict.GetM() || Ytest_predict.GetN() != NumClasses ) {
printf("Performance. Error: Size mismatch. Return NAN");
return NAN;
}
if (NumClasses <= 2) {
printf("Performance. Error: Not multiclass classification. Return NAN");
return NAN;
}
double perf = 0.0;
// Compute ytest_predict
DVector ytest_predict(n);
double *y = ytest_predict.GetPointer();
for (long i = 0; i < n; i++) {
DVector row(NumClasses);
Ytest_predict.GetRow(i, row);
long idx = -1;
row.Max(idx);
y[i] = (double)idx;
}
// Accuracy
double *y1 = ytest_truth.GetPointer();
double *y2 = ytest_predict.GetPointer();
for (long i = 0; i < n; i++) {
perf += ((int)y1[i])==((int)y2[i]) ? 1.0 : 0.0;
}
perf = perf/n * 100.0;
return perf;
}
//--------------------------------------------------------------------------
int main(int argc, char **argv) {
// Temporary variables
long int i, j, k, ii, idx = 1;
double ElapsedTime;
// Arguments
int NumThreads = atoi(argv[idx++]);
char *FileTrain = argv[idx++];
char *FileTest = argv[idx++];
int NumClasses = atoi(argv[idx++]);
int d = atoi(argv[idx++]);
int r = atoi(argv[idx++]);
int Num_lambda = atoi(argv[idx++]);
double *List_lambda = (double *)malloc(Num_lambda*sizeof(double));
for (ii = 0; ii < Num_lambda; ii++) {
List_lambda[ii] = atof(argv[idx++]);
}
int Num_sigma = atoi(argv[idx++]);
double *List_sigma = (double *)malloc(Num_sigma*sizeof(double));
for (i = 0; i < Num_sigma; i++) {
List_sigma[i] = atof(argv[idx++]);
}
int MAXIT = atoi(argv[idx++]);
double TOL = atof(argv[idx++]);
bool verbose = atoi(argv[idx++]);
// Threading
#ifdef USE_OPENBLAS
openblas_set_num_threads(NumThreads);
#elif USE_ESSL
#elif USE_OPENMP
omp_set_num_threads(NumThreads);
#else
NumThreads = 1; // To avoid compiler warining of unused variable
#endif
PREPARE_CLOCK(1);
START_CLOCK;
// Read in X = Xtrain (n*d), y = ytrain (n*1),
// and X0 = Xtest (m*d), y0 = ytest (m*1)
DPointArray Xtrain; // read all data points from train
DPointArray Xtest; // read all data points from test
DVector ytrain; // Training labels
DVector ytest; // Testing labels (ground truth)
DVector ytest_predict; // Predictions
if (ReadData(FileTrain, Xtrain, ytrain, d) == 0) {
return -1;
}
if (ReadData(FileTest, Xtest, ytest, d) == 0) {
return -1;
}
END_CLOCK;
ElapsedTime = ELAPSED_TIME;
printf("OneVsAll: time loading data = %g seconds\n", ElapsedTime); fflush(stdout);
// For multiclass classification, need to convert a single vector
// ytrain to a matrix Ytrain. The "predictions" are stored in the
// corresponding matrix Ytest_predict. The vector ytest_predict is
DMatrix Ytrain;
ConvertYtrain(ytrain, Ytrain, NumClasses);
int Seed = 0; // initialize seed as zero
// Loop over List_lambda
for (ii = 0; ii < Num_lambda; ii++) {
double lambda = List_lambda[ii];
// Loop over List_sigma
for (k = 0; k < Num_sigma; k++) {
double sigma = List_sigma[k];
// Seed the RNG
srandom(Seed);
START_CLOCK;
// Generate feature matrix Xdata_randbin given Xdata
vector< vector< pair<int,double> > > instances_old, instances_new;
long Xtrain_N = Xtrain.GetN();
for(i=0;i<Xtrain_N;i++){
instances_old.push_back(vector<pair<int,double> >());
for(j=0;j<d;j++){
int index = j+1;
double *myXtrain = Xtrain.GetPointer();
double myXtrain_feature = myXtrain[j*Xtrain_N+i];
if (myXtrain_feature != 0)
instances_old.back().push_back(pair<int,double>(index, myXtrain_feature));
}
}
long Xtest_N = Xtest.GetN();
for(i=0;i<Xtest_N;i++){
instances_old.push_back(vector<pair<int,double> >());
for(j=0;j<d;j++){
int index = j+1;
double *myXtest = Xtest.GetPointer();
double myXtest_feature = myXtest[j*Xtest_N+i];
if (myXtest_feature != 0)
instances_old.back().push_back(pair<int,double>(index, myXtest_feature));
}
}
END_CLOCK;
printf("Train. RandBin: Time (in seconds) for converting data format: %g\n", ELAPSED_TIME);fflush(stdout);
// add 0 feature for Enxu's code
START_CLOCK;
random_binning_feature(d+1, r, instances_old, instances_new, sigma);
END_CLOCK;
printf("Train. RandBin: Time (in seconds) for generating random binning features: %g\n", ELAPSED_TIME);fflush(stdout);
START_CLOCK;
SPointArray Xdata_randbin; // Generate random binning features
long int nnz = r*(Xtrain_N + Xtest_N);
long int dd = 0;
for(i = 0; i < instances_new.size(); i++){
if(dd < instances_new[i][r-1].first)
dd = instances_new[i][r-1].first;
}
Xdata_randbin.Init(Xtrain_N+Xtest_N, dd, nnz);
long int ind = 0;
long int *mystart = Xdata_randbin.GetPointerStart();
int *myidx = Xdata_randbin.GetPointerIdx();
double *myX = Xdata_randbin.GetPointerX();
for(i = 0; i < instances_new.size(); i++){
if (i == 0)
mystart[i] = 0;
else
mystart[i] = mystart[i-1] + r;
for(j = 0; j < instances_new[i].size(); j++){
myidx[ind] = instances_new[i][j].first-1;
myX[ind] = instances_new[i][j].second;
ind++;
}
}
mystart[i] = nnz; // mystart has a length N+1
// generate random binning features for Xtrain and Xtest
SPointArray Xtrain; // Training points
SPointArray Xtest; // Testing points
long Row_start = 0;
Xdata_randbin.GetSubset(Row_start, Xtrain_N,Xtrain);
Xdata_randbin.GetSubset(Xtrain_N,Xtest_N,Xtest);
Xdata_randbin.ReleaseAllMemory();
END_CLOCK;
printf("Train. RandBin: Time (in seconds) for converting data format back: %g\n", ELAPSED_TIME);fflush(stdout);
printf("OneVsAll: n train = %ld, m test = %ld, r = %d, D = %ld, Gamma = %f, num threads = %d\n", Xtrain_N, Xtest_N, r, dd, sigma, NumThreads); fflush(stdout);
// solve (Z'Z + lambdaI)w = Z'y, note that we never explicitly form
// Z'Z since Z is a large sparse matrix N*dd
START_CLOCK;
int m = Ytrain.GetN(); // number of classes
long N = Xtrain.GetN(); // number of training points
long NN = Xtest.GetN(); // number of training points
long M = Xtrain.GetD(); // dimension of randome binning features
DMatrix Ytest_predict(NN,m);
DMatrix W(M,m);
SPointArray EYE;
EYE.Init(M,M,M);
mystart = EYE.GetPointerStart();
myidx = EYE.GetPointerIdx();
myX = EYE.GetPointerX();
for(i=0;i<M;i++){
mystart[i] = i;
myidx[i] = i;
myX[i] = 1;
}
mystart[i] = M+1; // mystart has a length N+1
for (i = 0; i < m; i++) {
DVector w;
w.Init(M);
DVector ytrain, yy;
Ytrain.GetColumn(i, ytrain);
Xtrain.MatVec(ytrain, yy, TRANSPOSE);
double NormRHS = yy.Norm2();
PCG pcg_solver;
pcg_solver.Solve<SPointArray, SPointArray>(Xtrain, yy, w, EYE, MAXIT, TOL, 1);
if (verbose) {
int Iter = 0;
const double *ResHistory = pcg_solver.GetResHistory(Iter);
printf("RLCM::Train, PCG. iteration = %d, Relative residual = %g\n",
Iter, ResHistory[Iter-1]/NormRHS);fflush(stdout);
}
pcg_solver.GetSolution(w);
W.SetColumn(i, w);
}
END_CLOCK;
printf("Train. RandBin: Time (in seconds) for solving linear system solution: %g\n", ELAPSED_TIME);fflush(stdout);
// y = Xtest*W = z(x)'*w
START_CLOCK;
Xtest.MatMat(W,Ytest_predict,NORMAL,NORMAL);
double accuracy = Performance(ytest, Ytest_predict, NumClasses);
END_CLOCK;
ElapsedTime = ELAPSED_TIME;
printf("Test. RandBin: param = %g %g, perf = %g, time = %g\n", sigma, lambda, accuracy, ElapsedTime); fflush(stdout);
}// End loop over List_sigma
}// End loop over List_lambda
// Clean up
free(List_sigma);
free(List_lambda);
return 0;
}