int prepareGconst_GMM(GMM * gmm)
{
int m;
for (m = 0; m < gmm->nmix; m++) {
if (gmm->full == FA) {
gmm->gauss[m].gconst = cal_gconst(gmm->gauss[m].var, gmm->dim);
} else {
gmm->gauss[m].gconst = cal_gconstf(gmm->gauss[m].cov, gmm->dim);
}
if (gmm->gauss[m].gconst == 0) {
return -1;
}
}
return (0);
}
int load_GMM(GMM * gmm, FILE * fp)
{
int m, l;
freadf(gmm->weight, sizeof(*(gmm->weight)), gmm->nmix, fp);
for (m = 0; m < gmm->nmix; m++) {
freadf(gmm->gauss[m].mean, sizeof(*(gmm->gauss[m].mean)), gmm->dim, fp);
if (gmm->full != TR) {
freadf(gmm->gauss[m].var, sizeof(*(gmm->gauss[m].var)), gmm->dim, fp);
gmm->gauss[m].gconst = cal_gconst(gmm->gauss[m].var, gmm->dim);
} else {
for (l = 0; l < gmm->dim; l++) {
freadf(gmm->gauss[m].cov[l],
sizeof(*(gmm->gauss[m].cov[l])), gmm->dim, fp);
}
}
}
return (0);
}
int main(int argc, char **argv)
{
FILE *fp = stdin;
GMM gmm, tgmm, floor;
double E = DEF_E, V = DEF_V, W = DEF_W,
*dat, *pd, *cb, *icb, *logwgd, logb, *sum, *sum_m, **sum_v, diff, sum_w,
ave_logp0, ave_logp1, change = MAXVALUE, tmp1, tmp2;
int ispipe, l, ll, L = DEF_L, m, M = DEF_M, N, t, T = DEF_T, S =
DEF_S, full = FULL, n1, i, j, Imin = DEF_IMIN, Imax =
DEF_IMAX, *tindex, *cntcb;
if ((cmnd = strrchr(argv[0], '/')) == NULL)
cmnd = argv[0];
else
cmnd++;
/* -- Check options -- */
while (--argc)
if (**++argv == '-') {
switch (*(*argv + 1)) {
case 'h':
usage(0);
break;
case 'l':
L = atoi(*++argv);
--argc;
break;
case 'm':
M = atoi(*++argv);
--argc;
break;
case 't':
T = atoi(*++argv);
--argc;
break;
case 's':
S = atoi(*++argv);
--argc;
break;
case 'a':
Imin = atoi(*++argv);
--argc;
break;
case 'b':
Imax = atoi(*++argv);
--argc;
break;
case 'e':
E = atof(*++argv);
--argc;
break;
case 'v':
V = atof(*++argv);
--argc;
break;
case 'w':
W = atof(*++argv);
--argc;
break;
case 'f':
full = 1 - full;
break;
default:
fprintf(stderr, "%s: Illegal option \"%s\".\n", cmnd, *argv);
usage(1);
}
} else
fp = getfp(*argv, "rb");
/* -- Count number of training vectors -- */
if (T == -1) {
ispipe = fseek(fp, 0L, SEEK_END);
T = (int) (ftell(fp) / (double) L / (double) sizeof(float));
rewind(fp);
if (ispipe == -1) { /* training data is from standard input via pipe */
fprintf(stderr,
"\n %s (Error) -t option must be specified for the standard input via pipe.\n",
cmnd);
usage(1);
}
}
/* Memory allocation */
/* Training data */
dat = dgetmem(T * L);
/* for VQ */
N = 1;
while (N < M)
N *= 2;
cb = dgetmem(N * L);
icb = dgetmem(L);
tindex = (int *) getmem(T, sizeof(int));
cntcb = (int *) getmem(M, sizeof(int));
/* GMM */
gmm.weight = dgetmem(M);
gmm.gauss = (Gauss *) getmem(M, sizeof(Gauss));
for (m = 0; m < M; m++) {
gmm.gauss[m].mean = dgetmem(L);
gmm.gauss[m].var = dgetmem(L);
if (full == 1) {
gmm.gauss[m].cov = (double **) malloc(sizeof(double *) * L);
gmm.gauss[m].inv = (double **) malloc(sizeof(double *) * L);
for (l = 0; l < L; l++) {
gmm.gauss[m].cov[l] = dgetmem(L);
gmm.gauss[m].inv[l] = dgetmem(L);
}
}
}
if (full == 1) {
floor.gauss = (Gauss *) getmem(1, sizeof(Gauss));
floor.gauss[0].cov = (double **) malloc(sizeof(double *) * L);
for (l = 0; l < L; l++)
floor.gauss[0].cov[l] = dgetmem(L);
sum_m = dgetmem(L);
sum_v = (double **) malloc(sizeof(double *) * L);
}
/* temporary */
tgmm.weight = dgetmem(M);
tgmm.gauss = (Gauss *) getmem(M, sizeof(Gauss));
for (m = 0; m < M; m++) {
tgmm.gauss[m].mean = dgetmem(L);
tgmm.gauss[m].var = dgetmem(L);
if (full == 1) {
tgmm.gauss[m].cov = (double **) malloc(sizeof(double *) * L);
tgmm.gauss[m].inv = (double **) malloc(sizeof(double *) * L);
for (l = 0; l < L; l++) {
tgmm.gauss[m].cov[l] = dgetmem(L);
tgmm.gauss[m].inv[l] = dgetmem(L);
}
}
}
logwgd = dgetmem(M);
sum = dgetmem(M);
/* Read training data */
freadf(dat, sizeof(*dat), T * L, fp);
/* Initialization of GMM parameters */
/* LBG */
vaverage(dat, L, T, icb);
lbg(dat, L, T, icb, 1, cb, N, ITER, MINTRAIN, S, CENTUP, DELTA, END);
for (t = 0, pd = dat; t < T; t++, pd += L) {
tindex[t] = vq(pd, cb, L, M);
cntcb[tindex[t]]++;
}
for (m = 0; m < M; m++)
if (cntcb[m] == 0) {
fprintf(stderr, "Error: No data for mixture No.%d\n", m);
usage(1);
}
fprintf(stderr, "T = %d L = %d M = %d\n", T, L, M);
/* flooring value for weights */
W = 1.0 / (double) M *(double) W;
/* weights */
for (m = 0, sum_w = 0.0; m < M; m++) {
gmm.weight[m] = (double) cntcb[m] / (double) T;
if (gmm.weight[m] < W)
gmm.weight[m] = W;
sum_w += gmm.weight[m];
}
if (sum_w != 1.0)
for (m = 0; m < M; m++)
gmm.weight[m] /= sum_w;
/* mean */
for (m = 0, pd = cb; m < M; m++, pd += L)
movem(pd, gmm.gauss[m].mean, sizeof(double), L);
/* variance */
if (full != 1) {
for (t = 0, pd = dat; t < T; t++, pd += L)
for (l = 0; l < L; l++) {
diff = gmm.gauss[tindex[t]].mean[l] - pd[l];
gmm.gauss[tindex[t]].var[l] += sq(diff);
}
for (m = 0; m < M; m++)
for (l = 0; l < L; l++) {
gmm.gauss[m].var[l] /= (double) cntcb[m];
if (gmm.gauss[m].var[l] < V)
gmm.gauss[m].var[l] = V;
}
for (m = 0; m < M; m++)
gmm.gauss[m].gconst = cal_gconst(gmm.gauss[m].var, L);
}
/* full covariance */
else {
for (t = 0, pd = dat; t < T; t++, pd += L) {
for (l = 0; l < L; l++) {
for (i = 0; i <= l; i++) {
if (l == i) {
diff =
(gmm.gauss[tindex[t]].mean[l] -
pd[l]) * (gmm.gauss[tindex[t]].mean[i] - pd[i]);
floor.gauss[0].cov[l][i] += diff;
}
}
}
}
for (l = 0; l < L; l++) {
for (i = 0; i <= l; i++) {
if (l == i) {
floor.gauss[0].cov[l][i] /= T;
floor.gauss[0].cov[l][i] *= V;
}
}
}
for (t = 0, pd = dat; t < T; t++, pd += L) {
for (l = 0; l < L; l++) {
for (i = 0; i <= l; i++) {
diff =
(gmm.gauss[tindex[t]].mean[l] -
pd[l]) * (gmm.gauss[tindex[t]].mean[i] - pd[i]);
gmm.gauss[tindex[t]].cov[l][i] += diff;
}
}
}
for (m = 0; m < M; m++)
for (l = 0; l < L; l++)
for (i = 0; i <= l; i++) {
gmm.gauss[m].cov[l][i] /= (double) cntcb[m];
}
}
/* EM training of GMM parameters */
for (i = 0; (i <= Imax) && ((i <= Imin) || (fabs(change) > E)); i++) {
if (full != 1)
fillz_gmm(&tgmm, M, L);
else
fillz_gmmf(&tgmm, M, L);
fillz(sum, sizeof(double), M);
if (full != 1) {
for (m = 0; m < M; m++)
gmm.gauss[m].gconst = cal_gconst(gmm.gauss[m].var, L);
} else {
for (m = 0, n1 = 0; m < M; m++) {
gmm.gauss[m].gconst = cal_gconstf(gmm.gauss[m].cov, L);
if (gmm.gauss[m].gconst == 0) {
n1++;
for (l = 0; l < L; l++)
gmm.gauss[m].cov[l][l] += floor.gauss[0].cov[l][l];
gmm.gauss[m].gconst = cal_gconstf(gmm.gauss[m].cov, L);
}
if (gmm.gauss[m].gconst == 0) {
fprintf(stderr, "ERROR : Can't caluculate covdet");
exit(EXIT_FAILURE);
}
/* calculate inv */
cal_inv(gmm.gauss[m].cov, gmm.gauss[m].inv, L);
}
}
if (full == 1)
fprintf(stderr, "%d cov can't caluculate covdet\n", n1);
for (t = 0, ave_logp1 = 0.0, pd = dat; t < T; t++, pd += L) {
for (m = 0, logb = LZERO; m < M; m++) {
if (full != 1) {
logwgd[m] = log_wgd(&gmm, m, pd, L);
logb = log_add(logb, logwgd[m]);
}
/* full */
else {
logwgd[m] = log_wgdf(&gmm, m, pd, L);
logb = log_add(logb, logwgd[m]);
}
}
ave_logp1 += logb;
for (m = 0; m < M; m++) {
tmp1 = exp(logwgd[m] - logb);
sum[m] += tmp1;
for (l = 0; l < L; l++) {
tmp2 = tmp1 * pd[l];
tgmm.gauss[m].mean[l] += tmp2;
if (full != 1)
tgmm.gauss[m].var[l] += tmp2 * pd[l];
else {
for (j = 0; j <= l; j++) {
tgmm.gauss[m].cov[l][j] +=
tmp1 * (pd[l] - gmm.gauss[m].mean[l]) * (pd[j] -
gmm.
gauss[m].mean
[j]);
}
}
}
}
}
/* Output average log likelihood at each iteration */
ave_logp1 /= (double) T;
if (i == 1 && m == 1)
ave_logp0 = ave_logp1;
fprintf(stderr, "iter %3d : ", i);
fprintf(stderr, "ave_logprob = %g", ave_logp1);
if (i) {
change = ave_logp1 - ave_logp0;
fprintf(stderr, " change = %g", change);
}
fprintf(stderr, "\n");
ave_logp0 = ave_logp1;
/* Update perameters */
/* weights */
for (m = 0; m < M; m++)
gmm.weight[m] = sum[m] / (double) T;
/* mean, variance */
for (m = 0; m < M; m++) {
for (l = 0; l < L; l++)
gmm.gauss[m].mean[l] = tgmm.gauss[m].mean[l] / sum[m];
if (full != 1) {
for (l = 0; l < L; l++) {
gmm.gauss[m].var[l] =
tgmm.gauss[m].var[l] / sum[m] - sq(gmm.gauss[m].mean[l]);
if (gmm.gauss[m].var[l] < V)
gmm.gauss[m].var[l] = V;
}
}
/* full */
else {
for (l = 0; l < L; l++) {
for (j = 0; j <= l; j++) {
gmm.gauss[m].cov[l][j] = tgmm.gauss[m].cov[l][j] / sum[m];
}
}
}
}
}
/* Output GMM parameters */
fwritef(gmm.weight, sizeof(double), M, stdout);
if (full != 1) {
for (m = 0; m < M; m++) {
fwritef(gmm.gauss[m].mean, sizeof(double), L, stdout);
fwritef(gmm.gauss[m].var, sizeof(double), L, stdout);
}
} else {
for (m = 0; m < M; m++) {
fwritef(gmm.gauss[m].mean, sizeof(double), L, stdout);
for (i = 0; i < L; i++)
for (j = 0; j < i; j++)
gmm.gauss[m].cov[j][i] = gmm.gauss[m].cov[i][j];
for (l = 0; l < L; l++)
fwritef(gmm.gauss[m].cov[l], sizeof(double), L, stdout);
}
}
return (0);
}
