CCalcWmML::CCalcWmML(int L, int K, int b)
{
this->L = L;
this->K = K;
this->b = b;
wm = new double [K+1];
kernel = new double [L+1];
kernelTruncated = new double [L+1];
c = new double [L+1]; //c_m : the inner product of all L-merEsts for two L-mer with m mismatches
cTr = new double [L+1]; // c using truncated g.
h = new double [L+1]; // h_m : for inner product of all gapped-kmer counts .
calcwm();
calcKernel();
calcc();
}
void mexFunction( int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[] )
{
double *xd; /* input data, size Nd*n */
double *xs; /* SVM data, size Ns*n */
double *y; /* output data, size Nd*1 */
double *a; /* SVM coefficients, size Ns*1 */
double b; /* bias term, scalar */
long Nd; /* number of input data */
long Ns; /* number of SVM data */
long n; /* dimension of input space */
long d; /* number of output dimensions */
int *dy; /* dimensions of output vector*/
long i,j; /* counters */
kernel ker; /* kernel function */
mxArray *F; /* field of MATLAB struct */
if (nrhs!=2)
mexErrMsgTxt("Invalid number of input arguments.");
if (nlhs>2)
mexErrMsgTxt("Invalid number of output arguments.");
/* --- check input arguments --- */
if ( mxGetNumberOfElements(prhs[0]) != 1 ||
!mxIsStruct(prhs[0]) ||
getSvm(prhs[0],&a,&xs,&Ns,&n,&b,&ker) )
mexErrMsgTxt("Invalid first argument (support vector machine).");
if ( mxIsEmpty(prhs[1]) ||
!mxIsDouble(prhs[1]) ||
mxIsComplex(prhs[1]) )
mexErrMsgTxt("Invalid second argument (input data).");
if (nlhs > 1) {
if ( (F=mxGetField(prhs[0],0,FNAME_PROB)) == NULL ||
mxIsEmpty(F) ||
mxGetNumberOfDimensions(F) > 2 ||
!mxIsDouble(F) ||
mxIsComplex(F) ||
mxGetNumberOfElements(F) != 2 )
mexErrMsgTxt("SVM does not contain valid probability information.");
}
/* --- get input arguments --- */
d = mxGetNumberOfDimensions(prhs[1]); /* get true number of input dims */
dy = (int*)mxGetDimensions(prhs[1]);
while ( d>2 && dy[d-1]==1 )
d--;
if ( dy[d-1]==n ) /* compute number of output dims */
d--;
else if ( n>1 )
mexErrMsgTxt("Sizes of SVM and input data do not match.");
xd = mxGetPr(prhs[1]); /* get input data */
Nd = 1;
for (i=0; i<d; i++)
Nd *= dy[i];
/* --- compute SVM output --- */
plhs[0] = mxCreateNumericArray(d,dy,mxDOUBLE_CLASS,mxREAL);
y = mxGetPr(plhs[0]);
for (i=0; i<Nd; i++) {
y[i] = b;
for (j=0; j<Ns; j++)
y[i] += a[j]*calcKernel(xd+i,Nd,xs+j,Ns,n,&ker);
}
/* --- compute probabilities --- */
if (nlhs > 1) {
double *p; /* probability matrix */
double predict; /* temporary computation value */
double probA = (mxGetPr(mxGetField(prhs[0],0,FNAME_PROB)))[0];
double probB = (mxGetPr(mxGetField(prhs[0],0,FNAME_PROB)))[1];
plhs[1] = mxCreateNumericArray(d,dy,mxDOUBLE_CLASS,mxREAL);
p = mxGetPr(plhs[1]);
/* probability computation due to svm.cpp */
#define min_prob 1e-7
for (i=0; i<Nd; i++) {
predict = y[i]*probA + probB;
if (predict<0)
predict = 1.0/(1.0+exp(predict));
else
predict = 1.0-1.0/(1.0+exp(-predict));
if (y[i]<0) /* always use PREDICTED class */
predict = 1-predict;
p[i] = (predict > 1-min_prob) ? 1-min_prob : predict;
}
}
}
/**
* Computes Gram matrix of a specified kernel. Given a data matrix
* X (n x d), it returns Gram matrix K (n x n).
*
* @param kernelType 'linear' | 'poly' | 'rbf' | 'cosine'
*
* @param kernelParam -- | degree | sigma | --
*
* @param X a matrix with each row being a feature vector
* @return Gram matrix (n x n)
*
*/
Matrix& calcKernel(std::string kernelType,
double kernelParam, Matrix& X) {
return calcKernel(kernelType, kernelParam, X, X);
}
