double
loglinear_method::run(const snp_row &row1, const snp_row &row2, float *output)
{
arma::mat count = joint_count( row1, row2, get_data( )->phenotype, m_weight );
size_t num_samples = arma::accu( count );
set_num_ok_samples( num_samples );
if( arma::min( arma::min( count ) ) < METHOD_SMALLEST_CELL_SIZE_BINOMIAL )
{
return -9;
}
std::vector<log_double> likelihood( m_models.size( ), 0.0 );
std::vector<double> bic( m_models.size( ), 0.0 );
for(int i = 0; i < m_models.size( ); i++)
{
likelihood[ i ] = m_models[ i ]->prob( count );
bic[ i ] = -2.0 * likelihood[ i ].log_value( ) + m_models[ i ]->df( ) * log( num_samples );
}
unsigned int best_model = std::distance( bic.begin( ), std::min_element( bic.begin( ) + 1, bic.end( ) ) );
double LR = -2.0*(likelihood[ best_model ].log_value( ) - likelihood[ 0 ].log_value( ));
try
{
double p_value = 1.0 - chi_square_cdf( LR, m_models[ 0 ]->df( ) - m_models[ best_model ]->df( ) );
output[ 0 ] = p_value;
return p_value;
}
catch(bad_domain_value &e)
{
}
return -9;
}
lexicalCache PhraseDictionaryMultiModelCounts::CacheLexicalStatistics( const Phrase &phraseS, const Phrase &phraseT, AlignVector &alignment, const vector<lexicalTable*> &tables, bool is_input )
{
//do all the necessary lexical table lookups and get counts, but don't apply weights yet
Word null;
if (is_input) {
null.CreateFromString(Input, m_input, "NULL", false);
} else {
null.CreateFromString(Output, m_output, "NULL", false);
}
lexicalCache ret;
// all target words have to be explained
for(size_t ti=0; ti<alignment.size(); ti++) {
const set< size_t > & srcIndices = alignment[ ti ];
Word t_word = phraseT.GetWord(ti);
vector<lexicalPair> ti_vector;
if (srcIndices.empty()) {
// explain unaligned word by NULL
vector<float> joint_count (m_numModels);
vector<float> marginals (m_numModels);
FillLexicalCountsJoint(null, t_word, joint_count, tables);
FillLexicalCountsMarginal(null, marginals, tables);
ti_vector.push_back(make_pair(joint_count, marginals));
} else {
for (set< size_t >::const_iterator si(srcIndices.begin()); si != srcIndices.end(); ++si) {
Word s_word = phraseS.GetWord(*si);
vector<float> joint_count (m_numModels);
vector<float> marginals (m_numModels);
FillLexicalCountsJoint(s_word, t_word, joint_count, tables);
FillLexicalCountsMarginal(s_word, marginals, tables);
ti_vector.push_back(make_pair(joint_count, marginals));
}
}
ret.push_back(ti_vector);
}
return ret;
}
// get lexical probability for single word alignment pair
double PhraseDictionaryMultiModelCounts::GetLexicalProbability( Word &wordS, Word &wordT, const vector<lexicalTable*> &tables, vector<float> &multimodelweights ) const
{
vector<float> joint_count (m_numModels);
vector<float> marginals (m_numModels);
FillLexicalCountsJoint(wordS, wordT, joint_count, tables);
FillLexicalCountsMarginal(wordS, marginals, tables);
double lexProb = m_combineFunction(joint_count, marginals, multimodelweights);
return lexProb;
}
double
stagewise_method::run(const snp_row &row1, const snp_row &row2, float *output)
{
std::vector<log_double> likelihood( m_models.size( ), 0.0 );
arma::mat count;
float min_samples = 0.0;
unsigned int sample_threshold = METHOD_SMALLEST_CELL_SIZE_BINOMIAL;
if( m_model == "binomial" )
{
count = joint_count( row1, row2, get_data( )->phenotype, m_weight );
set_num_ok_samples( (size_t) arma::accu( count ) );
min_samples = arma::min( arma::min( count ) );
}
else if( m_model == "normal" )
{
count = joint_count_cont( row1, row2, get_data( )->phenotype, m_weight );
set_num_ok_samples( (size_t) arma::accu( count.col( 1 ) ) );
min_samples = arma::min( count.col( 1 ) );
sample_threshold = METHOD_SMALLEST_CELL_SIZE_NORMAL;
}
if( min_samples < sample_threshold )
{
return -9;
}
for(int i = 0; i < m_models.size( ); i++)
{
likelihood[ i ] = m_models[ i ]->prob( count );
}
for(int i = 1; i < m_models.size( ); i++)
{
double LR = -2.0*(likelihood[ i ].log_value( ) - likelihood[ 0 ].log_value( ));
try
{
output[ i - 1 ] = 1.0 - chi_square_cdf( LR, m_models[ 0 ]->df( ) - m_models[ i ]->df( ) );
}
catch(bad_domain_value &e)
{
}
}
return output[ 0 ];
}
