Problem with full covariance matrix

[Marcin Plonkowski]

In file gauden.c is function:

gauden_invert_variance_full(gauden_t *g)
{
    uint32 i, j, k;

    for (i = 0; i < g->n_mgau; i++) {
    for (j = 0; j < g->n_feat; j++) {
        for (k = 0; k < g->n_density; k++) {
        /* Yes, this works in-place (we hope) */
        if (invert(g->fullvar[i][j][k], g->fullvar[i][j][k],
               g->veclen[j]) != S3_SUCCESS) {
 ~~~~~~~~~~

 where method invert has 2 the same parameters g->fullvar[i][j][k].

 And in file matrix.c is function:

invert(float32 ** ainv, float32 ** a, int32 n)
{
char uplo;
float32 **tmp_a;
int32 info, nrhs, i;

/* Construct an identity matrix. */
memset(ainv[0], 0, sizeof(float32) * n * n);
    for (i = 0; i < n; i++)
    ainv[i][i] = 1.0;

~~~~~~~~~~

This causes, that method invert calculate inversion of identity matrix, instead of inversion of g->fullvar[i][j][k].

Has anyone seen this error?

[Nickolay V. Shmyrev]

This is a bug indeed which appeared during sphinxtrain refactoring when invert funciton from sphinxtrain was replaced by sphinxbase. I fixed it in trunk, thank you very much for report!

[Marcin Plonkowski]

Thank you.

I have yet to find one small mistake.
In file s3acc_io.c (sphinxtrain-5prealpha-win32), in function rdacc_den_full, should be exclamation mark before ck_readable method:

int
rdacc_den_full(...)
{
...
sprintf(fn, "%s/gauden_counts", dir);

if (!ck_readable(fn)) {
  //^
    E_ERROR("Unable to access %s\n", fn);
    return S3_ERROR;
}

[Nickolay V. Shmyrev]

Thanks, we fixed this in sources quite some time ago, I probably need to update win32 package, it is slightly outdated.