Example #1
0
inline void
Pei( DistMatrix<T,U,V>& P, Int n, T alpha )
{
#ifndef RELEASE
    CallStackEntry entry("MakeIdentity");
#endif
    P.ResizeTo( n, n );
    const Int localHeight = P.LocalHeight();
    const Int localWidth = P.LocalWidth();
    const Int colShift = P.ColShift();
    const Int rowShift = P.RowShift();
    const Int colStride = P.ColStride();
    const Int rowStride = P.RowStride();
    for( Int jLoc=0; jLoc<localWidth; ++jLoc )
    {
        const Int j = rowShift + jLoc*rowStride;
        for( Int iLoc=0; iLoc<localHeight; ++iLoc )
        {
            const Int i = colShift + iLoc*colStride;
            P.SetLocal( iLoc, jLoc, T(1) );
            if( i == j )
                P.UpdateLocal( iLoc, jLoc, alpha );
        }
    }
}
Example #2
0
inline void
Pei( DistMatrix<T,U,V>& P, Int n, T alpha )
{
    DEBUG_ONLY(CallStackEntry cse("Pei"))
    P.Resize( n, n );
    const Int localHeight = P.LocalHeight();
    const Int localWidth = P.LocalWidth();
    const Int colShift = P.ColShift();
    const Int rowShift = P.RowShift();
    const Int colStride = P.ColStride();
    const Int rowStride = P.RowStride();
    for( Int jLoc=0; jLoc<localWidth; ++jLoc )
    {
        const Int j = rowShift + jLoc*rowStride;
        for( Int iLoc=0; iLoc<localHeight; ++iLoc )
        {
            const Int i = colShift + iLoc*colStride;
            P.SetLocal( iLoc, jLoc, T(1) );
            if( i == j )
                P.UpdateLocal( iLoc, jLoc, alpha );
        }
    }
}
Example #3
0
int 
main( int argc, char* argv[] )
{
    Environment env( argc, argv );
    mpi::Comm comm = mpi::COMM_WORLD;

    try
    {
        const Int m = Input("--numExamples","number of examples",200);
        const Int n = Input("--numFeatures","number of features",100);
        const bool useIPM = Input("--useIPM","use Interior Point?",true);
        const Int maxIter = Input("--maxIter","maximum # of iter's",500);
        const Real gamma = Input("--gamma","two-norm coefficient",1.);
        const Real rho = Input("--rho","augmented Lagrangian param.",1.);
        const bool inv = Input("--inv","use explicit inverse",true);
        const bool progress = Input("--progress","print progress?",true);
        const bool display = Input("--display","display matrices?",false);
        const bool print = Input("--print","print matrices",false);
        ProcessInput();
        PrintInputReport();

        // Define a random (affine) hyperplane 
        DistMatrix<Real> w;
        Gaussian( w, n, 1 );
        {
            const Real wNorm = FrobeniusNorm( w );
            w *= 1/wNorm;
        }
        Real offset = SampleNormal();
        mpi::Broadcast( offset, 0, comm );

        // Draw each example (row) from a Gaussian perturbation of a point
        // lying on the hyperplane
        DistMatrix<Real> G;
        Gaussian( G, m, n );
        DistMatrix<Real,MR,STAR> w_MR_STAR;
        w_MR_STAR.AlignWith( G );
        w_MR_STAR = w;
        for( Int jLoc=0; jLoc<G.LocalWidth(); ++jLoc )
            for( Int iLoc=0; iLoc<G.LocalHeight(); ++iLoc )
                G.UpdateLocal( iLoc, jLoc, w_MR_STAR.GetLocal(jLoc,0)*offset );
        
        // Label each example based upon its location relative to the hyperplane
        DistMatrix<Real> q;
        Ones( q, m, 1 );
        Gemv( NORMAL, Real(1), G, w, -offset, q );
        auto sgnMap = []( Real alpha ) 
                      { return alpha >= 0 ? Real(1) : Real(-1); }; 
        EntrywiseMap( q, function<Real(Real)>(sgnMap) );

        if( mpi::Rank(comm) == 0 )
            Output("offset=",offset);
        if( print )
        {
            Print( w, "w" );
            Print( G, "G" );
            Print( q, "q" );
        }
        if( display )
            Display( G, "G" );

        SVMCtrl<Real> ctrl;
        ctrl.useIPM = useIPM;
        ctrl.modelFitCtrl.rho = rho;
        ctrl.modelFitCtrl.maxIter = maxIter;
        ctrl.modelFitCtrl.inv = inv;
        ctrl.modelFitCtrl.progress = progress;

        Timer timer;
        DistMatrix<Real> wHatSVM;
        if( mpi::Rank() == 0 )
            timer.Start();
        SVM( G, q, gamma, wHatSVM, ctrl );
        if( mpi::Rank() == 0 )
            timer.Stop();
        auto wSVM = View( wHatSVM, 0, 0, n, 1 );
        const Real offsetSVM = -wHatSVM.Get(n,0);
        const Real wSVMNorm = FrobeniusNorm( wSVM );
        if( mpi::Rank() == 0 )
        {
            Output("SVM time: ",timer.Total()," secs");
            Output
            ("|| wSVM ||_2=",wSVMNorm,"\n",
             "margin      =",Real(2)/wSVMNorm,"\n",
             "offsetSVM   =",offsetSVM,"\n",
             "offsetSVM / || wSVM ||_2=",offsetSVM/wSVMNorm);
        }
        if( print )
            Print( wSVM, "wSVM" );

        // Report the classification percentage using the returned hyperplane
        DistMatrix<Real> qSVM;
        Ones( qSVM, m, 1 );
        Gemv( NORMAL, Real(1), G, wSVM, -offsetSVM, qSVM );
        EntrywiseMap( qSVM, function<Real(Real)>(sgnMap) );
        if( print )
            Print( qSVM, "qSVM" );
        qSVM -= q;
        const Real numWrong = OneNorm(qSVM) / Real(2);
        if( mpi::Rank(comm) == 0 )
            Output("ratio misclassified: ",numWrong,"/",m);
    }
    catch( exception& e ) { ReportException(e); }

    return 0;
}
int 
main( int argc, char* argv[] )
{
    Initialize( argc, argv );

    try
    {
        const Int m = Input("--numExamples","number of examples",200);
        const Int n = Input("--numFeatures","number of features",100);
        const Int maxIter = Input("--maxIter","maximum # of iter's",500);
        const Real gamma = Input("--gamma","two-norm coefficient",1.);
        const Int penaltyInt = Input("--penalty","0: none, 1: l1, 2: l2",1);
        const Real rho = Input("--rho","augmented Lagrangian param.",1.);
        const bool inv = Input("--inv","use explicit inverse",true);
        const bool progress = Input("--progress","print progress?",true);
        const bool display = Input("--display","display matrices?",false);
        const bool print = Input("--print","print matrices",false);
        ProcessInput();
        PrintInputReport();

        Regularization penalty = static_cast<Regularization>(penaltyInt);

        // Define a random (affine) hyperplane 
        DistMatrix<Real> w;
        Gaussian( w, n, 1 );
        {
            const Real wNorm = FrobeniusNorm( w );
            Scale( Real(1)/wNorm, w );
        }
        Real offset = SampleNormal();
        mpi::Broadcast( offset, 0, mpi::COMM_WORLD );

        // Draw each example (row) from a Gaussian perturbation of a point
        // lying on the hyperplane
        DistMatrix<Real> G;
        Gaussian( G, m, n );
        DistMatrix<Real,MR,STAR> w_MR_STAR;
        w_MR_STAR.AlignWith( G );
        w_MR_STAR = w;
        for( Int jLoc=0; jLoc<G.LocalWidth(); ++jLoc )
            for( Int iLoc=0; iLoc<G.LocalHeight(); ++iLoc )
                G.UpdateLocal( iLoc, jLoc, w_MR_STAR.GetLocal(jLoc,0)*offset );
        
        // Label each example based upon its location relative to the hyperplane
        DistMatrix<Real> q;
        Ones( q, m, 1 );
        Gemv( NORMAL, Real(1), G, w, -offset, q );
        auto sgnMap = []( Real alpha ) 
                      { return alpha >= 0 ? Real(1) : Real(-1); }; 
        EntrywiseMap( q, function<Real(Real)>(sgnMap) );

        if( mpi::WorldRank() == 0 )
            cout << "offset=" << offset << endl;
        if( print )
        {
            Print( w, "w" );
            Print( G, "G" );
            Print( q, "q" );
        }
        if( display )
            Display( G, "G" );

        DistMatrix<Real> wHatLog;
        LogisticRegression
        ( G, q, wHatLog, gamma, penalty, rho, maxIter, inv, progress );
        auto wLog = View( wHatLog, 0, 0, n, 1 );
        const Real offsetLog = -wHatLog.Get(n,0);
        const Real wLogOneNorm = OneNorm( wLog );
        const Real wLogFrobNorm = FrobeniusNorm( wLog );
        if( mpi::WorldRank() == 0 )
            cout << "|| wLog ||_1=" << wLogOneNorm << "\n"
                 << "|| wLog ||_2=" << wLogFrobNorm << "\n"
                 << "margin      =" << Real(2)/wLogFrobNorm << "\n"
                 << "offsetLog=" << offsetLog << "\n"
                 << "offsetLog / || wLog ||_2=" << offsetLog/wLogFrobNorm 
                 << endl;
        if( print )
            Print( wLog, "wLog" );

        // Report the classification percentage using the returned hyperplane
        // TODO: Evaluate probabilities implied by logistic function
        DistMatrix<Real> qLog;
        Ones( qLog, m, 1 );
        Gemv( NORMAL, Real(1), G, wLog, -offsetLog, qLog );
        EntrywiseMap( qLog, function<Real(Real)>(sgnMap) );
        if( print )
            Print( qLog, "qLog" );
        Axpy( Real(-1), q, qLog );
        const Real numWrong = OneNorm(qLog) / Real(2);
        if( mpi::WorldRank() == 0 )
            cout << "ratio misclassified: " << numWrong << "/" << m << endl;
    }
    catch( exception& e ) { ReportException(e); }

    Finalize();
    return 0;
}