static void compute_sum_dcov_thread(int ni,int ki,int di, const float *v,const float *mu_old,const float *p, float *mu,float *sigma,float *w, int n_thread) { long n=ni,d=di,k=ki; compute_sum_dcov_t t= { n,k,d, v,mu_old,p, fvec_new(n_thread*d*k), /* mu */ fvec_new(n_thread*d*k), /* sigma */ fvec_new(n_thread*k), /* w */ n_thread }; compute_tasks(n_thread,n_thread,&compute_sum_dcov_task_fun,&t); /* accumulate over n's */ long i; fvec_cpy(mu,t.mu,k*d); fvec_cpy(sigma,t.sigma,k*d); fvec_cpy(w,t.w,k); for(i=1; i<n_thread; i++) { fvec_add(mu,t.mu+i*d*k,d*k); fvec_add(sigma,t.sigma+i*d*k,d*k); fvec_add(w,t.w+i*k,k); } free(t.mu); free(t.sigma); free(t.w); }
float fvec_median_const (const float *f, int n) { float *f2 = fvec_new (n); fvec_cpy (f2, f, n); float med = fvec_median (f2, n); free (f2); return med; }
void fvec_cpy_subvectors (const float * v, int * idx, int d, int nout, float * vout) { long i; for (i = 0 ; i < nout ; i++) fvec_cpy (vout + i * d, v + (long) idx[i] * d, d); }
/* estimate the GMM parameters */ static void gmm_compute_params (int n, const float * v, const float * p, gmm_t * g, int flags, int n_thread) { long i, j; long d=g->d, k=g->k; float * vtmp = fvec_new (d); float * mu_old = fvec_new_cpy (g->mu, k * d); float * w_old = fvec_new_cpy (g->w, k); fvec_0 (g->w, k); fvec_0 (g->mu, k * d); fvec_0 (g->sigma, k * d); if(0) { /* slow and simple */ for (j = 0 ; j < k ; j++) { double dtmp = 0; for (i = 0 ; i < n ; i++) { /* contribution to the gaussian weight */ dtmp += p[i * k + j]; /* contribution to mu */ fvec_cpy (vtmp, v + i * d, d); fvec_mul_by (vtmp, d, p[i * k + j]); fvec_add (g->mu + j * d, vtmp, d); /* contribution to the variance */ fvec_cpy (vtmp, v + i * d, d); fvec_sub (vtmp, mu_old + j * d, d); fvec_sqr (vtmp, d); fvec_mul_by (vtmp, d, p[i * k + j]); fvec_add (g->sigma + j * d, vtmp, d); } g->w[j] = dtmp; } } else { /* fast and complicated */ if(n_thread<=1) compute_sum_dcov(n,k,d,v,mu_old,p,g->mu,g->sigma,g->w); else compute_sum_dcov_thread(n,k,d,v,mu_old,p,g->mu,g->sigma,g->w,n_thread); } if(flags & GMM_FLAGS_1SIGMA) { for (j = 0 ; j < k ; j++) { float *sigma_j=g->sigma+j*d; double var=fvec_sum(sigma_j,d)/d; fvec_set(sigma_j,d,var); } } long nz=0; for(i=0; i<k*d; i++) if(g->sigma[i]<min_sigma) { g->sigma[i]=min_sigma; nz++; } if(nz) printf("WARN %ld sigma diagonals are too small (set to %g)\n",nz,min_sigma); for (j = 0 ; j < k ; j++) { fvec_div_by (g->mu + j * d, d, g->w[j]); fvec_div_by (g->sigma + j * d, d, g->w[j]); } assert(finite(fvec_sum(g->mu, k*d))); fvec_normalize (g->w, k, 1); printf ("w = "); fvec_print (g->w, k); double imfac = k * fvec_sum_sqr (g->w, k); printf (" imfac = %.3f\n", imfac); free (vtmp); free (w_old); free (mu_old); }
void vlad_compute(int k, int d, const float *centroids, int n, const float *v,int flags, float *desc) { int i,j,l,n_quantile,i0,i1,ai,a,ma,ni; int *perm ; float un , diff; float *tab,*u,*avg,*sum,*mom2,*dists; int *hist,*assign; if(flags<11 || flags>=13) { assign=ivec_new(n); nn(n,k,d,centroids,v,assign,NULL,NULL); if(flags==6 || flags==7) { n_quantile = flags==6 ? 3 : 1; fvec_0(desc,k*d*n_quantile); perm = ivec_new(n); tab = fvec_new(n); ivec_sort_index(assign,n,perm); i0=0; for(i=0;i<k;i++) { i1=i0; while(i1<n && assign[perm[i1]]==i) { i1++; } if(i1==i0) continue; for(j=0;j<d;j++) { for(l=i0;l<i1;l++) { tab[l-i0]=v[perm[l]*d+j]; } ni=i1-i0; fvec_sort(tab,ni); for(l=0;l<n_quantile;l++) { desc[(i*d+j)*n_quantile+l]=(tab[(l*ni+ni/2)/n_quantile]-centroids[i*d+j])*ni; } } i0=i1; } free(perm); free(tab); } else if(flags==5) { fvec_0(desc,k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { desc[assign[i]*d+j]+=v[i*d+j]; } } } else if(flags==8 || flags==9) { fvec_0(desc,k*d); u = fvec_new(d); for(i=0;i<n;i++) { fvec_cpy(u,v+i*d,d); fvec_sub(u,centroids+assign[i]*d,d); un=(float)sqrt(fvec_norm2sqr(u,d)); if(un==0) continue; if(flags==8) { fvec_div_by(u,d,un); } else if(flags==9) { fvec_div_by(u,d,sqrt(un)); } fvec_add(desc+assign[i]*d,u,d); } free(u); } else if(flags==10) { fvec_0(desc,k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { desc[assign[i]*d+j]+=v[i*d+j]; } } for(i=0;i<k;i++) { fvec_normalize(desc+i*d,d,2.0); } } else if(flags==13) { fvec_0(desc,k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { desc[assign[i]*d+j]+=(float)sqr(v[i*d+j]-centroids[assign[i]*d+j]); } } } else if(flags==14) { avg = fvec_new_0(k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { avg[assign[i]*d+j]+=v[i*d+j]-centroids[assign[i]*d+j]; } } hist=ivec_new_histogram(k,assign,n); for(i=0;i<k;i++) { if(hist[i]>0) { for(j=0;j<d;j++) { avg[i*d+j]/=hist[i]; } } } free(hist); fvec_0(desc,k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { desc[assign[i]*d+j]+=(float)(sqr(v[i*d+j]-centroids[assign[i]*d+j]-avg[assign[i]*d+j])); } } fvec_sqrt(desc,k*d); free(avg); } else if(flags==15) { fvec_0(desc,k*d*2); sum = desc; for(i=0;i<n;i++) { for(j=0;j<d;j++) { sum[assign[i]*d+j]+=v[i*d+j]-centroids[assign[i]*d+j]; } } hist = ivec_new_histogram(k,assign,n); mom2 = desc+k*d; for(i=0;i<n;i++) { ai=assign[i]; for(j=0;j<d;j++) { mom2[ai*d+j]+=(float)(sqr(v[i*d+j]-centroids[ai*d+j]-sum[ai*d+j]/hist[ai])); } } fvec_sqrt(mom2,k*d); free(hist); } else if(flags==17) { fvec_0(desc,k*d*2); for(i=0;i<n;i++) { for(j=0;j<d;j++) { diff=v[i*d+j]-centroids[assign[i]*d+j]; if(diff>0) { desc[assign[i]*d+j]+=diff; } else { desc[assign[i]*d+j+k*d]-=diff; } } } } else { fvec_0(desc,k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { desc[assign[i]*d+j]+=v[i*d+j]-centroids[assign[i]*d+j]; } } if(flags==1) { hist=ivec_new_histogram(k,assign,n); /* printf("unbalance factor=%g\n",ivec_unbalanced_factor(hist,k)); */ for(i=0;i<k;i++) { for(j=0;j<d;j++) { desc[i*d+j]/=hist[i]; } } free(hist); } if(flags==2) { for(i=0;i<k;i++) { fvec_normalize(desc+i*d,d,2.0); } } if(flags==3 || flags==4) { assert(!"not implemented"); } if(flags==16) { hist=ivec_new_histogram(k,assign,n); for(i=0;i<k;i++) { if(hist[i]>0) { fvec_norm(desc+i*d,d,2); fvec_mul_by(desc+i*d,d,sqrt(hist[i])); } } free(hist); } } free(assign); } else if(flags==11 || flags==12) { ma=flags==11 ? 4 : 2; assign=ivec_new(n*ma); dists=knn(n,k,d,ma,centroids,v,assign,NULL,NULL); fvec_0(desc,k*d); for(i=0;i<n;i++) { for(j=0;j<d;j++) { for(a=0;a<ma;a++) { desc[assign[ma*i+a]*d+j]+=v[i*d+j]-centroids[assign[ma*i+a]*d+j]; } } } free(dists); free(assign); } }
void gmm_fisher_spatial(int N, int K, int D, const float *Q, const float *sgmm, const float *ll, float *sdesc) { float *Q_sum = fvec_new_0(K); { long k, n; for(n = 0; n < N; n++) for(k = 0; k < K; k++) Q_sum[k] += Q[n * K + k]; for(k = 0; k < K; k++) Q_sum[k] /= N; } float *Q_ll, *Q_ll_2; { /* prepare a matrix containing both ll and ll**2 */ float *ll_ll2 = fvec_new(D * 2 * N); fvec_cpy(ll_ll2, ll, D * N); float *ll2 = ll_ll2 + D * N; long i; for(i = 0; i < D * N; i++) ll2[i] = ll[i] * ll[i]; /* compute Q.T * ll_ll2 */ FINTEGER mi = K, ni = 2 * D, ki = N; float one_over_N = 1.0 / N, zero = 0; Q_ll = fvec_new(K * 2 * D); Q_ll_2 = Q_ll + K * D; sgemm_("N", "N", &mi, &ni, &ki, &one_over_N, Q, &mi, ll_ll2, &ki, &zero, Q_ll, &mi); free(ll_ll2); } { const float *mm = sgmm; float *d_mm = sdesc; long k, d; for(d = 0; d < D; d++) for(k = 0; k < K; k++) d_mm[d + k * D] = Q_ll[K * d + k] - Q_sum[k] * mm[d]; float *d_S = sdesc + K * D; const float *S = sgmm + D; for(d = 0; d < D; d++) { float dfact = S[d] - mm[d] * mm[d]; for(k = 0; k < K; k++) d_S[d + k * D] = -Q_ll_2[K * d + k] + 2 * Q_ll[K * d + k] * mm[d] + Q_sum[k] * dfact; } } free(Q_ll); free(Q_sum); }