// Loads a new feature space and appends it to the multiset
void loadFeatureSpace( string filename, CDataSet< vSparseSample >& ds )
{
// Load the data
cout<<"Loading "<<filename<<"... "; cout.flush();
CDataSet<CSparseSample> d;
parseSparseFile( filename, d, ',', ',', '=' );
cout<<"Loaded "<<d.size()<<" samples; "<<nFeats( d )<<" features"<<endl;
// Append the Zhang dataset to the multiset
expand( ds, d );
cout<<" The combined dataset is now at "<<ds.size()<<" samples"<<endl;
}
int main( int argc, char* argv[] )
{
// Display usage and/or parse the parameters
if( argc < 2 ) { cout<<"Usage: "<<argv[0]<<" <options file>"<<endl; return -1; }
CGOStrutParams params; params.load( argv[1] ); params.display();
typedef CKernelLoss<CSparseSample> CSparseKLoss;
CDataSet<vSparseSample>::binop_t fcker = CCompositeSparseKernel( false );
CDataSet<CSparseSample>::binop_t fker = CSparseKernel( true );
CDataSet<CSparseSample>::binop_t floss = CSparseKLoss( fker );
function< double( double, double ) > fioker = CProdJointKernel();
///////////////////////////////
CDataSet< vSparseSample > dsi;
// Load the datasets
loadFeatureSpace( "mousefunc/ge_zhang.sdat.gz", dsi );
loadFeatureSpace( "mousefunc/ge_su.sdat.gz", dsi );
loadFeatureSpace( "mousefunc/i_adj.sdat", dsi );
loadFeatureSpace( "mousefunc/dd_pfam.sdat", dsi );
loadFeatureSpace( "mousefunc/dd_inter.sdat", dsi );
loadFeatureSpace( "mousefunc/phylo.sdat", dsi );
shared_ptr< CFeatMap > pfm( new CFeatMap );
CDataSet<CSparseSample> dsoTr;
CDataSet<CSparseSample> dsoTe;
// Load the appropriate set of labels
if( GO::hasMF( params.ontology() ) )
loadOntologySpace( "molecular function", "mf", dsoTr, dsoTe, pfm );
if( GO::hasBP( params.ontology() ) )
loadOntologySpace( "biological process", "bp", dsoTr, dsoTe, pfm );
if( GO::hasCC( params.ontology() ) )
loadOntologySpace( "cellular compoennt", "cc", dsoTr, dsoTe, pfm );
if( dsoTr.size() < 1 || dsoTe.size() < 1 )
throw std::runtime_error( "No labels loaded" );
// Append hierarchy to the samples
cout<<"Training data output space has "<<dsoTr.size()<<" samples; "<<nFeats( dsoTr )<<" features"<<endl;
cout<<"Test data output space has "<<dsoTe.size()<<" samples; "<<nFeats( dsoTe )<<" features"<<endl;
// DEBUG
/* vector< unsigned int > v;
for( unsigned int i = 0; i < 400; i++ )
v.push_back( i );
dsoTr.subsample( v );*/
// END OF DEBUG
// Combine the data to form a joint output space
typedef CIODataSet<vSparseSample, CSparseSample> CIOSet;
shared_ptr< CIOSet > pdsio( new CIOSet(fcker, fker, floss, fioker) );
pdsio->addSets( dsi, dsoTr );
pdsio->addSets( dsi, dsoTe );
cout<<"Input space has "<<pdsio->sizeI()<<" samples"<<endl;
cout<<"Output space has "<<pdsio->sizeO()<<" samples, "<<nFeats( pdsio->getO() )<<" features"<<endl;
string ont;
if( GO::hasMF( params.ontology() ) ) ont += "mf";
if( GO::hasBP( params.ontology() ) ) ont += "bp";
if( GO::hasCC( params.ontology() ) ) ont += "cc";
string fnScores = "mfunc-score-" + params.alg_choice() + "-" + ont + ".sdat";
cout<<"Saving prediction scores to "<<fnScores<<endl;
// Re-split the data into training and test sets, both using the same output space now
// Retrieve the fold sizes
vector< unsigned int > foldSizes;
foldSizes.push_back( pdsio->sizeI() - dsoTe.size() );
foldSizes.push_back( dsoTe.size() );
// Create the ranges
virange_t vTrain, vTest;
splitCV( foldSizes, 1, vTrain, vTest );
display( vTrain, vTest );
// Perform the split
shared_ptr< CIOSet > pTrain;
shared_ptr< CIOSet > pTest;
cout<<"About to split"<<endl;
pdsio->splitTrainTest( vTrain, vTest, pTrain, pTest );
cout<<"Finished the split"<<endl;
pTrain->cache();
pTest->cache();
// Create and add the appropriate test classifier
shared_ptr< CClassifier<vSparseSample, CSparseSample> > pclsf = createClassifier<vSparseSample, CSparseSample>( params );
// Train and test the classifier
pclsf->train( pTrain );
vector< double > loss = pclsf->test( pTest );
// Report the results
double m = std::accumulate( loss.begin(), loss.end(), 0.0 ) /
static_cast<double>( loss.size() );
cout<<"Mean loss per test sample: "<<m<<endl;
pTrain->cacheIExternal( pTest->getI() );
predScores( *pclsf, pTest->getI(), fnScores );
return 0;
}
