void add_cutting_plane(
bmrm_ll** tail,
bool* map,
float64_t* A,
uint32_t free_idx,
float64_t* cp_data,
uint32_t dim)
{
ASSERT(map[free_idx]);
LIBBMRM_MEMCPY(A+free_idx*dim, cp_data, dim*sizeof(float64_t));
map[free_idx]=false;
bmrm_ll *cp=(bmrm_ll*)LIBBMRM_CALLOC(1, sizeof(bmrm_ll));
if (cp==NULL)
{
SG_SERROR("Out of memory.\n");
return;
}
cp->address=A+(free_idx*dim);
cp->prev=*tail;
cp->next=NULL;
cp->idx=free_idx;
(*tail)->next=cp;
*tail=cp;
}
SGMatrix<float64_t> SGMatrix<T>::matrix_multiply(
SGMatrix<float64_t> A, SGMatrix<float64_t> B,
bool transpose_A, bool transpose_B, float64_t scale)
{
/* these variables store size of transposed matrices*/
index_t cols_A=transpose_A ? A.num_rows : A.num_cols;
index_t rows_A=transpose_A ? A.num_cols : A.num_rows;
index_t rows_B=transpose_B ? B.num_cols : B.num_rows;
index_t cols_B=transpose_B ? B.num_rows : B.num_cols;
/* do a dimension check */
if (cols_A!=rows_B)
{
SG_SERROR("SGMatrix::matrix_multiply(): Dimension mismatch: "
"A(%dx%d)*B(%dx%D)\n", rows_A, cols_A, rows_B, cols_B);
}
/* allocate result matrix */
SGMatrix<float64_t> C(rows_A, cols_B);
/* multiply */
cblas_dgemm(CblasColMajor,
transpose_A ? CblasTrans : CblasNoTrans,
transpose_B ? CblasTrans : CblasNoTrans,
rows_A, cols_B, cols_A, scale,
A.matrix, A.num_rows, B.matrix, B.num_rows,
0.0, C.matrix, C.num_rows);
return C;
}
CKLDualInferenceMethod* CKLDualInferenceMethod::obtain_from_generic(
CInference* inference)
{
if (inference==NULL)
return NULL;
if (inference->get_inference_type()!=INF_KL_DUAL)
{
SG_SERROR("Provided inference is not of type CKLDualInferenceMethod!\n");
}
SG_REF(inference);
return (CKLDualInferenceMethod*)inference;
}
SGVector<float64_t> SGMatrix<T>::compute_eigenvectors(SGMatrix<float64_t> matrix)
{
if (matrix.num_rows!=matrix.num_rows)
{
SG_SERROR("SGMatrix::compute_eigenvectors(SGMatrix<float64_t>): matrix"
" rows and columns are not equal!\n");
}
/* use reference counting for SGVector */
SGVector<float64_t> result(NULL, 0, true);
result.vlen=matrix.num_rows;
result.vector=compute_eigenvectors(matrix.matrix, matrix.num_rows,
matrix.num_rows);
return result;
}
SEXP Rsg(SEXP args)
{
/* The SEXP (Simple Expression) args is a list of arguments of the .External call.
* it consists of "sg", "func" and additional arguments.
* */
try
{
if (!interface)
{
// init_shogun has to be called before anything else
// exit_shogun is called upon destruction of the interface (see
// destructor of CRInterface
init_shogun(&r_print_message, &r_print_warning,
&r_print_error, &r_cancel_computations);
interface=new CRInterface(args);
#ifdef HAVE_PYTHON
CPythonInterface::run_python_init();
#endif
#ifdef HAVE_OCTAVE
COctaveInterface::run_octave_init();
#endif
}
else
((CRInterface*) interface)->reset(args);
if (!interface->handle())
SG_SERROR("Unknown command.\n");
}
catch (std::bad_alloc)
{
error("Out of memory error.\n");
return R_NilValue;
}
catch (ShogunException e)
{
error("%s", e.get_exception_string());
return R_NilValue;
}
catch (...)
{
error("%s", "Returning from SHOGUN in error.");
return R_NilValue;
}
return ((CRInterface*) interface)->get_return_values();
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
try
{
if (!interface)
{
// init_shogun has to be called before anything else
// exit_shogun is called upon destruction of the interface (see
// destructor of CMatlabInterface
init_shogun(&matlab_print_message, &matlab_print_warning,
&matlab_print_error, &matlab_cancel_computations);
interface=new CMatlabInterface(nlhs, plhs, nrhs, prhs);
#ifdef HAVE_PYTHON
CPythonInterface::run_python_init();
#endif
#ifdef HAVE_OCTAVE
COctaveInterface::run_octave_init();
#endif
#ifdef HAVE_R
CRInterface::run_r_init();
#endif
}
else
((CMatlabInterface*) interface)->reset(nlhs, plhs, nrhs, prhs);
if (!interface->handle())
SG_SERROR("Unknown command.\n");
}
catch (std::bad_alloc)
{
mexErrMsgTxt("Out of memory error.");
}
catch (ShogunException e)
{
mexErrMsgTxt(e.get_exception_string());
}
catch (...)
{
mexErrMsgTxt("Returning from SHOGUN in error.");
}
}
double* SGMatrix<T>::compute_eigenvectors(double* matrix, int n, int m)
{
ASSERT(n == m);
char V='V';
char U='U';
int info;
int ord=n;
int lda=n;
double* eigenvalues=SG_CALLOC(float64_t, n+1);
// lapack sym matrix eigenvalues+vectors
wrap_dsyev(V, U, ord, matrix, lda,
eigenvalues, &info);
if (info!=0)
SG_SERROR("DSYEV failed with code %d\n", info);
return eigenvalues;
}
bmrm_return_value_T svm_bmrm_solver(
bmrm_data_T* data,
float64_t* W,
float64_t TolRel,
float64_t TolAbs,
float64_t lambda,
uint32_t _BufSize,
bool cleanICP,
uint32_t cleanAfter,
float64_t K,
uint32_t Tmax,
CRiskFunction* risk_function)
{
bmrm_return_value_T bmrm = {0, 0, 0, 0, 0, 0, 0};
libqp_state_T qp_exitflag;
float64_t *b, *beta, *diag_H, sq_norm_W;
float64_t R, *subgrad, *A, QPSolverTolRel, rsum, C=1.0;
uint32_t *I, *ICPcounter, *ICPs, cntICP=0;
uint8_t S = 1;
uint32_t nDim=data->w_dim;
CTime ttime;
float64_t tstart, tstop;
float64_t *b2, *beta2, *diag_H2, *A2, *H2;
uint32_t *I2, *ICPcounter2, nCP_new=0, idx=0, idx2=0, icp_iter=0, icp_iter2=0;
int32_t idx_icp=0, idx_icp2=0;
bool flag1=true, flag2=true;
tstart=ttime.cur_time_diff(false);
BufSize=_BufSize;
QPSolverTolRel=TolRel*0.5;
H=NULL;
b=NULL;
beta=NULL;
A=NULL;
subgrad=NULL;
diag_H=NULL;
I=NULL;
ICPcounter=NULL;
ICPs=NULL;
b2=NULL;
beta2=NULL;
H2=NULL;
I2=NULL;
ICPcounter2=NULL;
A2=NULL;
diag_H2=NULL;
H=(float64_t*)LIBBMRM_CALLOC(BufSize*BufSize, sizeof(float64_t));
if (H==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
A=(float64_t*)LIBBMRM_CALLOC(nDim*BufSize, sizeof(float64_t));
if (A==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
b=(float64_t*)LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
if (b==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
beta=(float64_t*)LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
if (beta==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
subgrad=(float64_t*)LIBBMRM_CALLOC(nDim, sizeof(float64_t));
if (subgrad==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
diag_H=(float64_t*)LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
if (diag_H==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
I=(uint32_t*)LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
if (I==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
ICPcounter=(uint32_t*)LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
if (ICPcounter==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
ICPs=(uint32_t*)LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
if (ICPs==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
/* Temporary buffers for ICP removal */
b2=(float64_t*)LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
beta2=(float64_t*)LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
ICPcounter2=(uint32_t*)LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
I2=(uint32_t*)LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
diag_H2=(float64_t*)LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
A2=(float64_t*)LIBBMRM_CALLOC(nDim*BufSize, sizeof(float64_t));
H2=(float64_t*)LIBBMRM_CALLOC(BufSize*BufSize, sizeof(float64_t));
if (b2==NULL || beta2==NULL || ICPcounter2==NULL || I2==NULL || diag_H2==NULL || A2==NULL || H2==NULL)
{
bmrm.exitflag=-2;
}
/* Iinitial solution */
risk_function->risk(data, &R, subgrad, W);
bmrm.nCP=0;
bmrm.nIter=0;
bmrm.exitflag=0;
b[0]=-R;
LIBBMRM_MEMCPY(A, subgrad, nDim*sizeof(float64_t));
/* Compute initial value of Fp, Fd, assuming that W is zero vector */
sq_norm_W=0;
bmrm.Fp=R+0.5*lambda*sq_norm_W;
bmrm.Fd=-LIBBMRM_PLUS_INF;
tstop=ttime.cur_time_diff(false);
/* Verbose output */
SG_SPRINT("%4d: tim=%.3lf, Fp=%lf, Fd=%lf, R=%lf\n",
bmrm.nIter, tstop-tstart, bmrm.Fp, bmrm.Fd, R);
/* main loop */
while (bmrm.exitflag==0)
{
tstart=ttime.cur_time_diff(false);
bmrm.nIter++;
/* Update H */
if (bmrm.nCP>0)
{
for (uint32_t i=0; i<bmrm.nCP; ++i)
{
rsum=0.0;
for (uint32_t j=0; j<nDim; ++j)
{
rsum+=A[LIBBMRM_INDEX(j, i, nDim)]*A[LIBBMRM_INDEX(j, bmrm.nCP, nDim)];
}
H[LIBBMRM_INDEX(i, bmrm.nCP, BufSize)]=rsum/lambda;
}
for (uint32_t i=0; i<bmrm.nCP; ++i)
{
H[LIBBMRM_INDEX(bmrm.nCP, i, BufSize)]=H[LIBBMRM_INDEX(i, bmrm.nCP, BufSize)];
}
}
H[LIBBMRM_INDEX(bmrm.nCP, bmrm.nCP, BufSize)]=0.0;
for (uint32_t i=0; i<nDim; ++i)
H[LIBBMRM_INDEX(bmrm.nCP, bmrm.nCP, BufSize)]+=A[LIBBMRM_INDEX(i, bmrm.nCP, nDim)]*A[LIBBMRM_INDEX(i, bmrm.nCP, nDim)]/lambda;
diag_H[bmrm.nCP]=H[LIBBMRM_INDEX(bmrm.nCP, bmrm.nCP, BufSize)];
I[bmrm.nCP]=1;
bmrm.nCP++;
/* call QP solver */
qp_exitflag = libqp_splx_solver(&get_col, diag_H, b, &C, I, &S, beta,
bmrm.nCP, QPSolverMaxIter, 0.0, QPSolverTolRel, -LIBBMRM_PLUS_INF, 0);
bmrm.qp_exitflag=qp_exitflag.exitflag;
/* Update ICPcounter (add one to unused and reset used) + compute number of active CPs*/
bmrm.nzA=0;
for (uint32_t aaa=0; aaa<bmrm.nCP; ++aaa)
{
if (beta[aaa]>epsilon)
{
bmrm.nzA+=1;
ICPcounter[aaa]=0;
} else {
ICPcounter[aaa]+=1;
}
}
/* W update */
for (uint32_t i=0; i<nDim; ++i)
{
rsum = 0.0;
for (uint32_t j=0; j<bmrm.nCP; ++j)
{
rsum += A[LIBBMRM_INDEX(i, j, nDim)]*beta[j];
}
W[i] = -rsum/lambda;
}
/* risk and subgradient computation */
risk_function->risk(data, &R, subgrad, W);
LIBBMRM_MEMCPY(A+bmrm.nCP*nDim, subgrad, nDim*sizeof(float64_t));
b[bmrm.nCP]=-R;
for (uint32_t j=0; j<nDim; ++j)
b[bmrm.nCP]+=subgrad[j]*W[j];
sq_norm_W = 0;
for (uint32_t j=0; j<nDim; ++j)
sq_norm_W+=W[j]*W[j];
bmrm.Fp=R+0.5*lambda*sq_norm_W;
bmrm.Fd=-qp_exitflag.QP;
/* Stopping conditions */
if (bmrm.Fp - bmrm.Fd <= TolRel*LIBBMRM_ABS(bmrm.Fp)) bmrm.exitflag=1;
if (bmrm.Fp - bmrm.Fd <= TolAbs) bmrm.exitflag=2;
if (bmrm.nCP >= BufSize) bmrm.exitflag=-1;
tstop=ttime.cur_time_diff(false);
/* Verbose output */
SG_SPRINT("%4d: tim=%.3lf, Fp=%lf, Fd=%lf, (Fp-Fd)=%lf, (Fp-Fd)/Fp=%lf, R=%lf, nCP=%d, nzA=%d\n",
bmrm.nIter, tstop-tstart, bmrm.Fp, bmrm.Fd, bmrm.Fp-bmrm.Fd, (bmrm.Fp-bmrm.Fd)/bmrm.Fp, R, bmrm.nCP, bmrm.nzA);
/* Check size of Buffer */
if (bmrm.nCP>=BufSize)
{
bmrm.exitflag=-2;
SG_SERROR("Buffer exceeded.\n");
}
/* Inactive Cutting Planes (ICP) removal */
if (cleanICP)
{
/* find ICP */
cntICP = 0;
for (uint32_t aaa=0; aaa<bmrm.nCP; ++aaa)
if (ICPcounter[aaa]>=cleanAfter)
{
ICPs[cntICP++]=aaa;
}
/* do ICP */
if (cntICP > 0)
{
nCP_new=bmrm.nCP-cntICP;
idx=0;
idx2=0;
icp_iter=0;
icp_iter2=0;
idx_icp=ICPs[icp_iter];
idx_icp2=ICPs[icp_iter2];
flag1=true; flag2=true;
for (uint32_t i=0; i<bmrm.nCP; ++i)
{
if ((int32_t)i != idx_icp)
{
b2[idx]=b[i];
beta2[idx]=beta[i];
ICPcounter2[idx]=ICPcounter[i];
I2[idx]=I[i];
diag_H2[idx]=diag_H[i];
LIBBMRM_MEMCPY(A2+idx*nDim, A+i*nDim, nDim*sizeof(float64_t));
idx2=0;
icp_iter2=0;
idx_icp2=ICPs[icp_iter2];
flag2=true;
for (uint32_t j=0; j<bmrm.nCP; ++j)
{
if ((int32_t)j != idx_icp2)
{
H2[LIBBMRM_INDEX(idx, idx2, BufSize)]=H[LIBBMRM_INDEX(i, j, BufSize)];
idx2++;
} else {
if (flag2 && icp_iter2+1 < cntICP)
{
idx_icp2=ICPs[++icp_iter2];
} else {
flag2=false;
idx_icp2=-1;
}
}
}
idx++;
} else {
if (flag1 && icp_iter+1 < cntICP)
{
idx_icp=ICPs[++icp_iter];
} else {
flag1=false;
idx_icp=-1;
}
}
}
/* copy data from tmps back to original */
LIBBMRM_MEMCPY(b, b2, nCP_new*sizeof(float64_t));
LIBBMRM_MEMCPY(beta, beta2, nCP_new*sizeof(float64_t));
LIBBMRM_MEMCPY(ICPcounter, ICPcounter2, nCP_new*sizeof(uint32_t));
LIBBMRM_MEMCPY(I, I2, nCP_new*sizeof(uint32_t));
LIBBMRM_MEMCPY(diag_H, diag_H2, nCP_new*sizeof(float64_t));
LIBBMRM_MEMCPY(A, A2, nDim*nCP_new*sizeof(float64_t));
for (uint32_t i=0; i<nCP_new; ++i)
for (uint32_t j=0; j<nCP_new; ++j)
H[LIBBMRM_INDEX(i, j, BufSize)]=H2[LIBBMRM_INDEX(i, j, BufSize)];
bmrm.nCP=nCP_new;
}
}
} /* end of main loop */
cleanup:
LIBBMRM_FREE(H);
LIBBMRM_FREE(b);
LIBBMRM_FREE(beta);
LIBBMRM_FREE(A);
LIBBMRM_FREE(subgrad);
LIBBMRM_FREE(diag_H);
LIBBMRM_FREE(I);
LIBBMRM_FREE(ICPcounter);
LIBBMRM_FREE(ICPs);
LIBBMRM_FREE(H2);
LIBBMRM_FREE(b2);
LIBBMRM_FREE(beta2);
LIBBMRM_FREE(A2);
LIBBMRM_FREE(diag_H2);
LIBBMRM_FREE(I2);
LIBBMRM_FREE(ICPcounter2);
return(bmrm);
}
SparsityStructure* CSparseMatrixOperator<complex128_t>
::get_sparsity_structure(int64_t power) const
{
SG_SERROR("Not supported for complex128_t\n");
return new SparsityStructure();
}
bmrm_return_value_T svm_bmrm_solver(
CStructuredModel* model,
float64_t* W,
float64_t TolRel,
float64_t TolAbs,
float64_t _lambda,
uint32_t _BufSize,
bool cleanICP,
uint32_t cleanAfter,
float64_t K,
uint32_t Tmax,
bool verbose)
{
bmrm_return_value_T bmrm;
libqp_state_T qp_exitflag={0, 0, 0, 0};
float64_t *b, *beta, *diag_H, *prevW;
float64_t R, *subgrad, *A, QPSolverTolRel, C=1.0, wdist=0.0;
floatmax_t rsum, sq_norm_W, sq_norm_Wdiff=0.0;
uint32_t *I;
uint8_t S=1;
uint32_t nDim=model->get_dim();
CTime ttime;
float64_t tstart, tstop;
bmrm_ll *CPList_head, *CPList_tail, *cp_ptr, *cp_ptr2, *cp_list=NULL;
float64_t *A_1=NULL, *A_2=NULL;
bool *map=NULL;
tstart=ttime.cur_time_diff(false);
BufSize=_BufSize;
QPSolverTolRel=1e-9;
H=NULL;
b=NULL;
beta=NULL;
A=NULL;
subgrad=NULL;
diag_H=NULL;
I=NULL;
prevW=NULL;
H= (float64_t*) LIBBMRM_CALLOC(BufSize*BufSize, sizeof(float64_t));
if (H==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
A= (float64_t*) LIBBMRM_CALLOC(nDim*BufSize, sizeof(float64_t));
if (A==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
b= (float64_t*) LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
if (b==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
beta= (float64_t*) LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
if (beta==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
subgrad= (float64_t*) LIBBMRM_CALLOC(nDim, sizeof(float64_t));
if (subgrad==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
diag_H= (float64_t*) LIBBMRM_CALLOC(BufSize, sizeof(float64_t));
if (diag_H==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
I= (uint32_t*) LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
if (I==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
ICP_stats icp_stats;
icp_stats.maxCPs = BufSize;
icp_stats.ICPcounter= (uint32_t*) LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
if (icp_stats.ICPcounter==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
icp_stats.ICPs= (float64_t**) LIBBMRM_CALLOC(BufSize, sizeof(float64_t*));
if (icp_stats.ICPs==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
icp_stats.ACPs= (uint32_t*) LIBBMRM_CALLOC(BufSize, sizeof(uint32_t));
if (icp_stats.ACPs==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
/* Temporary buffers for ICP removal */
icp_stats.H_buff= (float64_t*) LIBBMRM_CALLOC(BufSize*BufSize, sizeof(float64_t));
if (icp_stats.H_buff==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
map= (bool*) LIBBMRM_CALLOC(BufSize, sizeof(bool));
if (map==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
memset( (bool*) map, true, BufSize);
cp_list= (bmrm_ll*) LIBBMRM_CALLOC(1, sizeof(bmrm_ll));
if (cp_list==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
prevW= (float64_t*) LIBBMRM_CALLOC(nDim, sizeof(float64_t));
if (prevW==NULL)
{
bmrm.exitflag=-2;
goto cleanup;
}
bmrm.hist_Fp = SGVector< float64_t >(BufSize);
bmrm.hist_Fd = SGVector< float64_t >(BufSize);
bmrm.hist_wdist = SGVector< float64_t >(BufSize);
/* Iinitial solution */
R=model->risk(subgrad, W);
bmrm.nCP=0;
bmrm.nIter=0;
bmrm.exitflag=0;
b[0]=-R;
/* Cutting plane auxiliary double linked list */
LIBBMRM_MEMCPY(A, subgrad, nDim*sizeof(float64_t));
map[0]=false;
cp_list->address=&A[0];
cp_list->idx=0;
cp_list->prev=NULL;
cp_list->next=NULL;
CPList_head=cp_list;
CPList_tail=cp_list;
/* Compute initial value of Fp, Fd, assuming that W is zero vector */
sq_norm_W=0;
bmrm.Fp=R+0.5*_lambda*sq_norm_W;
bmrm.Fd=-LIBBMRM_PLUS_INF;
tstop=ttime.cur_time_diff(false);
/* Verbose output */
if (verbose)
SG_SPRINT("%4d: tim=%.3lf, Fp=%lf, Fd=%lf, R=%lf\n",
bmrm.nIter, tstop-tstart, bmrm.Fp, bmrm.Fd, R);
/* store Fp, Fd and wdist history */
bmrm.hist_Fp[0]=bmrm.Fp;
bmrm.hist_Fd[0]=bmrm.Fd;
bmrm.hist_wdist[0]=0.0;
/* main loop */
while (bmrm.exitflag==0)
{
tstart=ttime.cur_time_diff(false);
bmrm.nIter++;
/* Update H */
if (bmrm.nCP>0)
{
A_2=get_cutting_plane(CPList_tail);
cp_ptr=CPList_head;
for (uint32_t i=0; i<bmrm.nCP; ++i)
{
A_1=get_cutting_plane(cp_ptr);
cp_ptr=cp_ptr->next;
rsum= SGVector<float64_t>::dot(A_1, A_2, nDim);
H[LIBBMRM_INDEX(bmrm.nCP, i, BufSize)]
= H[LIBBMRM_INDEX(i, bmrm.nCP, BufSize)]
= rsum/_lambda;
}
}
A_2=get_cutting_plane(CPList_tail);
rsum = SGVector<float64_t>::dot(A_2, A_2, nDim);
H[LIBBMRM_INDEX(bmrm.nCP, bmrm.nCP, BufSize)]=rsum/_lambda;
diag_H[bmrm.nCP]=H[LIBBMRM_INDEX(bmrm.nCP, bmrm.nCP, BufSize)];
I[bmrm.nCP]=1;
bmrm.nCP++;
beta[bmrm.nCP]=0.0; // [beta; 0]
#if 0
/* TODO: scaling...*/
float64_t scale = SGVector<float64_t>::max(diag_H, BufSize)/(1000.0*_lambda);
SGVector<float64_t> sb(bmrm.nCP);
sb.zero();
sb.vec1_plus_scalar_times_vec2(sb.vector, 1/scale, b, bmrm.nCP);
SGVector<float64_t> sh(bmrm.nCP);
sh.zero();
sb.vec1_plus_scalar_times_vec2(sh.vector, 1/scale, diag_H, bmrm.nCP);
qp_exitflag =
libqp_splx_solver(&get_col, sh.vector, sb.vector, &C, I, &S, beta,
bmrm.nCP, QPSolverMaxIter, 0.0, QPSolverTolRel, -LIBBMRM_PLUS_INF, 0);
#else
/* call QP solver */
qp_exitflag=libqp_splx_solver(&get_col, diag_H, b, &C, I, &S, beta,
bmrm.nCP, QPSolverMaxIter, 0.0, QPSolverTolRel, -LIBBMRM_PLUS_INF, 0);
#endif
bmrm.qp_exitflag=qp_exitflag.exitflag;
/* Update ICPcounter (add one to unused and reset used)
* + compute number of active CPs */
bmrm.nzA=0;
for (uint32_t aaa=0; aaa<bmrm.nCP; ++aaa)
{
if (beta[aaa]>epsilon)
{
++bmrm.nzA;
icp_stats.ICPcounter[aaa]=0;
}
else
{
icp_stats.ICPcounter[aaa]+=1;
}
}
/* W update */
memset(W, 0, sizeof(float64_t)*nDim);
cp_ptr=CPList_head;
for (uint32_t j=0; j<bmrm.nCP; ++j)
{
A_1=get_cutting_plane(cp_ptr);
cp_ptr=cp_ptr->next;
SGVector<float64_t>::vec1_plus_scalar_times_vec2(W, -beta[j]/_lambda, A_1, nDim);
}
/* risk and subgradient computation */
R = model->risk(subgrad, W);
add_cutting_plane(&CPList_tail, map, A,
find_free_idx(map, BufSize), subgrad, nDim);
sq_norm_W=SGVector<float64_t>::dot(W, W, nDim);
b[bmrm.nCP]=SGVector<float64_t>::dot(subgrad, W, nDim) - R;
sq_norm_Wdiff=0.0;
for (uint32_t j=0; j<nDim; ++j)
{
sq_norm_Wdiff+=(W[j]-prevW[j])*(W[j]-prevW[j]);
}
bmrm.Fp=R+0.5*_lambda*sq_norm_W;
bmrm.Fd=-qp_exitflag.QP;
wdist=CMath::sqrt(sq_norm_Wdiff);
/* Stopping conditions */
if (bmrm.Fp - bmrm.Fd <= TolRel*LIBBMRM_ABS(bmrm.Fp))
bmrm.exitflag=1;
if (bmrm.Fp - bmrm.Fd <= TolAbs)
bmrm.exitflag=2;
if (bmrm.nCP >= BufSize)
bmrm.exitflag=-1;
tstop=ttime.cur_time_diff(false);
/* Verbose output */
if (verbose)
SG_SPRINT("%4d: tim=%.3lf, Fp=%lf, Fd=%lf, (Fp-Fd)=%lf, (Fp-Fd)/Fp=%lf, R=%lf, nCP=%d, nzA=%d, QPexitflag=%d\n",
bmrm.nIter, tstop-tstart, bmrm.Fp, bmrm.Fd, bmrm.Fp-bmrm.Fd,
(bmrm.Fp-bmrm.Fd)/bmrm.Fp, R, bmrm.nCP, bmrm.nzA, qp_exitflag.exitflag);
/* Keep Fp, Fd and w_dist history */
bmrm.hist_Fp[bmrm.nIter]=bmrm.Fp;
bmrm.hist_Fd[bmrm.nIter]=bmrm.Fd;
bmrm.hist_wdist[bmrm.nIter]=wdist;
/* Check size of Buffer */
if (bmrm.nCP>=BufSize)
{
bmrm.exitflag=-2;
SG_SERROR("Buffer exceeded.\n");
}
/* keep W (for wdist history track) */
LIBBMRM_MEMCPY(prevW, W, nDim*sizeof(float64_t));
/* Inactive Cutting Planes (ICP) removal */
if (cleanICP)
{
clean_icp(&icp_stats, bmrm, &CPList_head, &CPList_tail, H, diag_H, beta, map, cleanAfter, b, I);
}
} /* end of main loop */
bmrm.hist_Fp.resize_vector(bmrm.nIter);
bmrm.hist_Fd.resize_vector(bmrm.nIter);
bmrm.hist_wdist.resize_vector(bmrm.nIter);
cp_ptr=CPList_head;
while(cp_ptr!=NULL)
{
cp_ptr2=cp_ptr;
cp_ptr=cp_ptr->next;
LIBBMRM_FREE(cp_ptr2);
cp_ptr2=NULL;
}
cp_list=NULL;
cleanup:
LIBBMRM_FREE(H);
LIBBMRM_FREE(b);
LIBBMRM_FREE(beta);
LIBBMRM_FREE(A);
LIBBMRM_FREE(subgrad);
LIBBMRM_FREE(diag_H);
LIBBMRM_FREE(I);
LIBBMRM_FREE(icp_stats.ICPcounter);
LIBBMRM_FREE(icp_stats.ICPs);
LIBBMRM_FREE(icp_stats.ACPs);
LIBBMRM_FREE(icp_stats.H_buff);
LIBBMRM_FREE(map);
LIBBMRM_FREE(prevW);
if (cp_list)
LIBBMRM_FREE(cp_list);
return(bmrm);
}
complex128_t SGMatrix<complex128_t>::max_single()
{
SG_SERROR("SGMatrix::max_single():: Not supported for complex128_t\n");
return complex128_t(0.0);
}
void SGSparseMatrix<complex64_t>::load(CFile* loader)
{
SG_SERROR("SGSparseMatrix::load():: Not supported for complex64_t");
}
template<class T> CRegressionLabels* SGSparseMatrix<T>::load_svmlight_file(char* fname,
bool do_sort_features)
{
CRegressionLabels* lab=NULL;
size_t blocksize=1024*1024;
size_t required_blocksize=blocksize;
uint8_t* dummy=SG_MALLOC(uint8_t, blocksize);
FILE* f=fopen(fname, "ro");
if (f)
{
free_data();
SG_SINFO("counting line numbers in file %s\n", fname)
size_t sz=blocksize;
size_t block_offs=0;
size_t old_block_offs=0;
fseek(f, 0, SEEK_END);
size_t fsize=ftell(f);
rewind(f);
while (sz == blocksize)
{
sz=fread(dummy, sizeof(uint8_t), blocksize, f);
for (size_t i=0; i<sz; i++)
{
block_offs++;
if (dummy[i]=='\n' || (i==sz-1 && sz<blocksize))
{
num_vectors++;
required_blocksize=CMath::max(required_blocksize, block_offs-old_block_offs+1);
old_block_offs=block_offs;
}
}
SG_SPROGRESS(block_offs, 0, fsize, 1, "COUNTING:\t")
}
SG_SINFO("found %d feature vectors\n", num_vectors)
SG_FREE(dummy);
blocksize=required_blocksize;
dummy = SG_MALLOC(uint8_t, blocksize+1); //allow setting of '\0' at EOL
lab=new CRegressionLabels(num_vectors);
sparse_matrix=SG_MALLOC(SGSparseVector<T>, num_vectors);
rewind(f);
sz=blocksize;
int32_t lines=0;
while (sz == blocksize)
{
sz=fread(dummy, sizeof(uint8_t), blocksize, f);
size_t old_sz=0;
for (size_t i=0; i<sz; i++)
{
if (i==sz-1 && dummy[i]!='\n' && sz==blocksize)
{
size_t len=i-old_sz+1;
uint8_t* data=&dummy[old_sz];
for (size_t j=0; j<len; j++)
dummy[j]=data[j];
sz=fread(dummy+len, sizeof(uint8_t), blocksize-len, f);
i=0;
old_sz=0;
sz+=len;
}
if (dummy[i]=='\n' || (i==sz-1 && sz<blocksize))
{
size_t len=i-old_sz;
uint8_t* data=&dummy[old_sz];
int32_t dims=0;
for (size_t j=0; j<len; j++)
{
if (data[j]==':')
dims++;
}
if (dims<=0)
{
SG_SERROR("Error in line %d - number of"
" dimensions is %d line is %d characters"
" long\n line_content:'%.*s'\n", lines,
dims, len, len, (const char*) data);
}
SGSparseVectorEntry<T>* feat=SG_MALLOC(SGSparseVectorEntry<T>, dims);
size_t j=0;
for (; j<len; j++)
{
if (data[j]==' ')
{
data[j]='\0';
lab->set_label(lines, atof((const char*) data));
break;
}
}
int32_t d=0;
j++;
uint8_t* start=&data[j];
for (; j<len; j++)
{
if (data[j]==':')
{
data[j]='\0';
feat[d].feat_index=(int32_t) atoi((const char*) start)-1;
num_features=CMath::max(num_features, feat[d].feat_index+1);
j++;
start=&data[j];
for (; j<len; j++)
{
if (data[j]==' ' || data[j]=='\n')
{
data[j]='\0';
feat[d].entry=(T) atof((const char*) start);
d++;
break;
}
}
if (j==len)
{
data[j]='\0';
feat[dims-1].entry=(T) atof((const char*) start);
}
j++;
start=&data[j];
}
}
sparse_matrix[lines].num_feat_entries=dims;
sparse_matrix[lines].features=feat;
old_sz=i+1;
lines++;
SG_SPROGRESS(lines, 0, num_vectors, 1, "LOADING:\t")
}
}
}
SG_SINFO("file successfully read\n")
fclose(f);
}
void SGSparseMatrix<complex64_t>::save(CFile* saver)
{
SG_SERROR("SGSparseMatrix::save():: Not supported for complex64_t");
}
SGVector<index_t> SGVector<complex64_t>::argsort()
{
SG_SERROR("SGVector::argsort():: Not supported for complex64_t\n");
SGVector<index_t> idx(vlen);
return idx;
}
void arpack_dsaupd(double* matrix, int n, int nev, const char* which,
int mode, bool pos, double shift, double* eigenvalues,
double* eigenvectors, int& status)
{
// check if nev is greater than n
if (nev>n)
SG_SERROR("Number of required eigenpairs is greater than order of the matrix");
// check specified mode
if (mode!=1 && mode!=3)
SG_SERROR("Unknown mode specified");
// init ARPACK's reverse communication parameter
// (should be zero initially)
int ido = 0;
// specify that non-general eigenproblem will be solved
// (Ax=lGx, where G=I)
char bmat[2] = "I";
// init tolerance (zero means machine precision)
double tol = 0.0;
// allocate array to hold residuals
double* resid = new double[n];
// set number of Lanczos basis vectors to be used
// (with max(4*nev,n) sufficient for most tasks)
int ncv = nev*4>n ? n : nev*4;
// allocate array 'v' for dsaupd routine usage
int ldv = n;
double* v = new double[ldv*ncv];
// init array for i/o params for routine
int* iparam = new int[11];
// specify method for selecting implicit shifts (1 - exact shifts)
iparam[0] = 1;
// specify max number of iterations
iparam[2] = 2*2*n;
// set the computation mode (1 for regular or 3 for shift-inverse)
iparam[6] = mode;
// init array indicating locations of vectors for routine callback
int* ipntr = new int[11];
// allocate workaround arrays
double* workd = new double[3*n];
int lworkl = ncv*(ncv+8);
double* workl = new double[lworkl];
// init info holding status (should be zero at first call)
int info = 0;
// which eigenpairs to find
char* which_ = strdup(which);
// All
char* all_ = strdup("A");
// shift-invert mode
if (mode==3)
{
for (int i=0; i<n; i++)
matrix[i*n+i] -= shift;
if (pos)
{
clapack_dpotrf(CblasColMajor,CblasUpper,n,matrix,n);
clapack_dpotri(CblasColMajor,CblasUpper,n,matrix,n);
}
else
{
int* ipiv = new int[n];
clapack_dgetrf(CblasColMajor,n,n,matrix,n,ipiv);
clapack_dgetri(CblasColMajor,n,matrix,n,ipiv);
delete[] ipiv;
}
}
// main computation loop
do
{
dsaupd_(&ido, bmat, &n, which_, &nev, &tol, resid,
&ncv, v, &ldv, iparam, ipntr, workd, workl,
&lworkl, &info);
if ((ido==1)||(ido==-1))
{
cblas_dsymv(CblasColMajor,CblasUpper,
n,1.0,matrix,n,
(workd+ipntr[0]-1),1,
0.0,(workd+ipntr[1]-1),1);
}
} while ((ido==1)||(ido==-1));
// check if DSAUPD failed
if (info<0)
{
if ((info<=-1)&&(info>=-6))
SG_SWARNING("DSAUPD failed. Wrong parameter passed.");
else if (info==-7)
SG_SWARNING("DSAUPD failed. Workaround array size is not sufficient.");
else
SG_SWARNING("DSAUPD failed. Error code: %d.", info);
status = -1;
}
else
{
if (info==1)
SG_SWARNING("Maximum number of iterations reached.\n");
// allocate select for dseupd
int* select = new int[ncv];
// allocate d to hold eigenvalues
double* d = new double[2*ncv];
// sigma for dseupd
double sigma = shift;
// init ierr indicating dseupd possible errors
int ierr = 0;
// specify that eigenvectors to be computed too
int rvec = 1;
dseupd_(&rvec, all_, select, d, v, &ldv, &sigma, bmat,
&n, which_, &nev, &tol, resid, &ncv, v, &ldv,
iparam, ipntr, workd, workl, &lworkl, &ierr);
if (ierr!=0)
{
SG_SWARNING("DSEUPD failed with status=%d", ierr);
status = -1;
}
else
{
for (int i=0; i<nev; i++)
{
eigenvalues[i] = d[i];
for (int j=0; j<n; j++)
eigenvectors[j*nev+i] = v[i*n+j];
}
}
// cleanup
delete[] select;
delete[] d;
}
// cleanup
delete[] all_;
delete[] which_;
delete[] resid;
delete[] v;
delete[] iparam;
delete[] ipntr;
delete[] workd;
delete[] workl;
};
