[176] | 1 | #include <vector> |
---|
[384] | 2 | #include "mixtures.h" |
---|
[176] | 3 | |
---|
[477] | 4 | namespace bdm { |
---|
[176] | 5 | |
---|
| 6 | |
---|
| 7 | void MixEF::init ( BMEF* Com0, const mat &Data, int c ) { |
---|
| 8 | //prepare sizes |
---|
| 9 | Coms.set_size ( c ); |
---|
[477] | 10 | n = c; |
---|
[180] | 11 | weights.set_parameters ( ones ( c ) ); //assume at least one observation in each comp. |
---|
[176] | 12 | //est will be done at the end |
---|
| 13 | // |
---|
| 14 | int i; |
---|
| 15 | int ndat = Data.cols(); |
---|
| 16 | //Estimate Com0 from all data |
---|
[286] | 17 | Coms ( 0 ) = Com0->_copy_(); |
---|
[176] | 18 | // Coms(0)->set_evalll(false); |
---|
| 19 | Coms ( 0 )->bayesB ( Data ); |
---|
| 20 | // Flatten it to its original shape |
---|
| 21 | Coms ( 0 )->flatten ( Com0 ); |
---|
| 22 | |
---|
[189] | 23 | //Copy it to the rest |
---|
[477] | 24 | for ( i = 1; i < n; i++ ) { |
---|
[176] | 25 | //copy Com0 and create new rvs for them |
---|
[286] | 26 | Coms ( i ) = Coms ( 0 )->_copy_ ( ); |
---|
[176] | 27 | } |
---|
| 28 | //Pick some data for each component and update it |
---|
[477] | 29 | for ( i = 0; i < n; i++ ) { |
---|
[176] | 30 | //pick one datum |
---|
[477] | 31 | int ind = floor ( ndat * UniRNG.sample() ); |
---|
| 32 | Coms ( i )->bayes ( Data.get_col ( ind ), 1.0 ); |
---|
[197] | 33 | //flatten back to oringinal |
---|
[477] | 34 | Coms ( i )->flatten ( Com0 ); |
---|
[176] | 35 | } |
---|
| 36 | |
---|
| 37 | //est already exists - must be deleted before build_est() can be used |
---|
| 38 | delete est; |
---|
| 39 | build_est(); |
---|
| 40 | |
---|
| 41 | } |
---|
| 42 | |
---|
[189] | 43 | void MixEF::bayesB ( const mat &data , const vec &wData ) { |
---|
[477] | 44 | int ndat = data.cols(); |
---|
| 45 | int t, i, niter; |
---|
| 46 | bool converged = false; |
---|
[176] | 47 | |
---|
| 48 | multiBM weights0 ( weights ); |
---|
| 49 | |
---|
| 50 | Array<BMEF*> Coms0 ( n ); |
---|
[477] | 51 | for ( i = 0; i < n; i++ ) { |
---|
| 52 | Coms0 ( i ) = ( BMEF* ) Coms ( i )->_copy_(); |
---|
| 53 | } |
---|
[176] | 54 | |
---|
[477] | 55 | niter = 0; |
---|
| 56 | mat W = ones ( n, ndat ) / n; |
---|
| 57 | mat Wlast = ones ( n, ndat ) / n; |
---|
[176] | 58 | vec w ( n ); |
---|
| 59 | vec ll ( n ); |
---|
| 60 | // tmp for weights |
---|
| 61 | vec wtmp = zeros ( n ); |
---|
[189] | 62 | int maxi; |
---|
| 63 | double maxll; |
---|
[176] | 64 | //Estim |
---|
| 65 | while ( !converged ) { |
---|
| 66 | // Copy components back to their initial values |
---|
| 67 | // All necessary information is now in w and Coms0. |
---|
| 68 | Wlast = W; |
---|
| 69 | // |
---|
[311] | 70 | //#pragma omp parallel for |
---|
[477] | 71 | for ( t = 0; t < ndat; t++ ) { |
---|
[311] | 72 | //#pragma omp parallel for |
---|
[477] | 73 | for ( i = 0; i < n; i++ ) { |
---|
| 74 | ll ( i ) = Coms ( i )->logpred ( data.get_col ( t ) ); |
---|
| 75 | wtmp = 0.0; |
---|
| 76 | wtmp ( i ) = 1.0; |
---|
[176] | 77 | ll ( i ) += weights.logpred ( wtmp ); |
---|
| 78 | } |
---|
[477] | 79 | |
---|
| 80 | maxll = max ( ll, maxi ); |
---|
| 81 | switch ( method ) { |
---|
| 82 | case QB: |
---|
| 83 | w = exp ( ll - maxll ); |
---|
| 84 | w /= sum ( w ); |
---|
| 85 | break; |
---|
| 86 | case EM: |
---|
| 87 | w = 0.0; |
---|
| 88 | w ( maxi ) = 1.0; |
---|
| 89 | break; |
---|
[189] | 90 | } |
---|
[477] | 91 | |
---|
[189] | 92 | W.set_col ( t, w ); |
---|
[176] | 93 | } |
---|
| 94 | |
---|
[189] | 95 | // copy initial statistics |
---|
[311] | 96 | //#pragma omp parallel for |
---|
[477] | 97 | for ( i = 0; i < n; i++ ) { |
---|
[176] | 98 | Coms ( i )-> set_statistics ( Coms0 ( i ) ); |
---|
| 99 | } |
---|
| 100 | weights.set_statistics ( &weights0 ); |
---|
| 101 | |
---|
[189] | 102 | // Update statistics |
---|
| 103 | // !!!! note wData ==> this is extra weight of the data record |
---|
| 104 | // !!!! For typical cases wData=1. |
---|
[477] | 105 | for ( t = 0; t < ndat; t++ ) { |
---|
[311] | 106 | //#pragma omp parallel for |
---|
[477] | 107 | for ( i = 0; i < n; i++ ) { |
---|
| 108 | Coms ( i )-> bayes ( data.get_col ( t ), W ( i, t ) * wData ( t ) ); |
---|
[176] | 109 | } |
---|
[189] | 110 | weights.bayes ( W.get_col ( t ) * wData ( t ) ); |
---|
[176] | 111 | } |
---|
| 112 | |
---|
| 113 | niter++; |
---|
| 114 | //TODO better convergence rule. |
---|
[477] | 115 | converged = ( niter > 10 );//( sumsum ( abs ( W-Wlast ) ) /n<0.1 ); |
---|
[176] | 116 | } |
---|
| 117 | |
---|
| 118 | //Clean Coms0 |
---|
[477] | 119 | for ( i = 0; i < n; i++ ) { |
---|
| 120 | delete Coms0 ( i ); |
---|
| 121 | } |
---|
[189] | 122 | } |
---|
| 123 | |
---|
| 124 | void MixEF::bayes ( const vec &data ) { |
---|
| 125 | |
---|
[176] | 126 | }; |
---|
| 127 | |
---|
[189] | 128 | void MixEF::bayes ( const mat &data ) { |
---|
| 129 | this->bayesB ( data, ones ( data.cols() ) ); |
---|
| 130 | }; |
---|
[176] | 131 | |
---|
[189] | 132 | |
---|
[176] | 133 | double MixEF::logpred ( const vec &dt ) const { |
---|
| 134 | |
---|
[477] | 135 | vec w = weights.posterior().mean(); |
---|
| 136 | double exLL = 0.0; |
---|
| 137 | for ( int i = 0; i < n; i++ ) { |
---|
| 138 | exLL += w ( i ) * exp ( Coms ( i )->logpred ( dt ) ); |
---|
[176] | 139 | } |
---|
| 140 | return log ( exLL ); |
---|
| 141 | } |
---|
[180] | 142 | |
---|
[477] | 143 | emix* MixEF::epredictor ( ) const { |
---|
[504] | 144 | Array<shared_ptr<epdf> > pC ( n ); |
---|
[477] | 145 | for ( int i = 0; i < n; i++ ) { |
---|
| 146 | pC ( i ) = Coms ( i )->epredictor ( ); |
---|
| 147 | } |
---|
[389] | 148 | emix* tmp; |
---|
[477] | 149 | tmp = new emix( ); |
---|
[504] | 150 | tmp->set_parameters ( weights.posterior().mean(), pC ); |
---|
[180] | 151 | return tmp; |
---|
| 152 | } |
---|
[197] | 153 | |
---|
[477] | 154 | void MixEF::flatten ( const BMEF* M2 ) { |
---|
| 155 | const MixEF* Mix2 = dynamic_cast<const MixEF*> ( M2 ); |
---|
[565] | 156 | bdm_assert_debug ( Mix2->n == n, "Different no of coms" ); |
---|
[197] | 157 | //Flatten each component |
---|
[477] | 158 | for ( int i = 0; i < n; i++ ) { |
---|
| 159 | Coms ( i )->flatten ( Mix2->Coms ( i ) ); |
---|
[197] | 160 | } |
---|
[198] | 161 | //Flatten weights = make them equal!! |
---|
[477] | 162 | weights.set_statistics ( & ( Mix2->weights ) ); |
---|
[254] | 163 | } |
---|
[271] | 164 | } |
---|