int main(int argc, char** argv) {
// seed random generator
srand(time(NULL));
ifstream infile("../iris.data");
string line;
DataVec data;
double fac = 4.0;
while (std::getline(infile, line))
{
std::vector<std::string> fields;
boost::split(fields,line, boost::is_any_of(","));
assert(fields.size() == 5);
double x1 = atof(fields[0].c_str())*fac;
double x2 = atof(fields[1].c_str())*fac;
double x3 = atof(fields[2].c_str())*fac;
double x4 = atof(fields[3].c_str())*fac;
Point p(x1,x2,x3,x4);
//std::cout << p << std::endl;
data.push_back(p);
}
std::cout << "Collected " << data.size() << " points. " << std::endl;
assert(data.size()>K);
// init centroids to random points
DataVec centroids;
//centroids.reserve(K);
DataVec::iterator dataBegin = random_unique(data.begin(),data.end(),K);
std::cout << K << " random points " << std::endl;
for (size_t i=0;i<K;++i) {
std::cout << data[i] << "\n";
centroids.push_back(data[i]);
}
std::cout << centroids.size() << " centroids " << std::endl;
// Lloyd's algorithm to iteratively fit the cluster centroids.
bool done = fit(data,centroids);
while(!done) {
done = fit(data,centroids);
std::cout << done << "\n";
}
double idx = dunnIndex(data,centroids);
cout << "Dunn Index for this clustering " << idx << "\n";
// write clustering to file
ofstream of("clusters.dat");
for (DataVec::iterator it = data.begin(); it!=data.end(); ++it) {
of << *it << std::endl;
}
of.close();
return EXIT_SUCCESS;
}
int main (int argc, char *argv[]) {
// handle cmd args
int batch_size, maxiter;
std::string datadir;
std::string output_file;
if ( argc > 5 || argc < 2 ) {
printf( "Usage: ./logistic_mpi <data_directory> <batch_size> "
"<max_iterations> <model_output_file>\n");
MPI_Finalize();
exit( 0 );
} else if ( argc == 5 ) {
datadir = argv[1];
batch_size = atoi( argv[2] ); // mini-batch processing
if ( batch_size == -1 ) { batch_size = INT_MIN; }
maxiter = atoi( argv[3] );
output_file = argv[4];
} else if ( argc == 4 ) {
datadir = argv[1];
batch_size = atoi( argv[2] ); // mini-batch processing
if ( batch_size == -1 ) { batch_size = INT_MIN; }
maxiter = atoi( argv[3] );
output_file = "logistic.model";
} else if ( argc == 4 ) {
datadir = argv[1];
batch_size = atoi( argv[2] ); // mini-batch processing
if ( batch_size == -1 ) { batch_size = INT_MIN; }
maxiter = 100;
output_file = "logistic.model";
} else {
datadir = argv[1];
batch_size = INT_MIN; // batch processing
maxiter = 100;
output_file = "logistic.model";
}
// initialize/populate mpi specific vars local to each node
double t1,t2; // elapsed time computation
int numtasks, taskid, len;
char hostname[MPI_MAX_PROCESSOR_NAME];
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numtasks);
MPI_Comm_rank(MPI_COMM_WORLD,&taskid);
MPI_Get_processor_name(hostname, &len);
MPI_Op op;
/* DATA PREPROCESSING */
if ( taskid == MASTER ) {
printf( "\nLoading and Preprocessing Data\n" );
}
t1 = MPI_Wtime();
// determine number of instances
DataVec datavec;
mlu::count_instances( datadir, datavec, num_inst );
// determine number of features
mlu::count_features( datavec[0], n );
/* DATA INITIALIZATION */
// randomize instances
std::random_shuffle( datavec.begin(), datavec.end() );
// partition data based on taskid
size_t div = datavec.size() / numtasks;
ProbSize limit = ( taskid == numtasks - 1 ) ? num_inst : div * ( taskid + 1 );
m = limit - div * taskid;
// danamically allocate data
Mat X( m, n );
Vec labels( m );
// load data partition
double feat_val, label;
ProbSize i = 0;
for ( ProbSize idx = taskid * div; idx < limit; ++idx ) {
std::ifstream data( datavec[idx] );
for ( ProbSize j=0; j<n; ++j ) {
data >> feat_val;
X(i,j) = feat_val;
}
data >> label;
labels[i] = label;
i++;
}
// perform feature scaling (optional)
if ( scaling ) {
// Allreduce to find global min
Vec X_min_tmp = X.colwise().minCoeff();
X_min_data = X_min_tmp.data();
Vec X_min = Vec( X_min_tmp.size() );
MPI_Allreduce( X_min_tmp.data(), X_min.data(), X_min_tmp.size(), MPI_DOUBLE, MPI_MIN, MPI_COMM_WORLD );
// Allreduce to find global max
Vec X_max_tmp = X.colwise().maxCoeff();
X_max_data = X_max_tmp.data();
Vec X_max = Vec( X_max_tmp.size() );
MPI_Allreduce( X_max_tmp.data(), X_max.data(), X_max_tmp.size(), MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD );
// scale features using global min and max
mlu::scale_features( X, X_min, X_max, 1, 0 );
}
/* FORMAT LABELS */
// get unique labels
mlu::get_unique_labels( labels, classmap );
// allreduce to obtain maximum label set size
int local_size = classmap.size();
int max_size = 0;
MPI_Allreduce( &local_size, &max_size, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD );
// allreduce to obtain global unique label set
int unique_labels[max_size];
int global_unique_labels[max_size];
for ( int i=0; i<max_size; ++i ) {
unique_labels[i] = -1;
global_unique_labels[i] = -1;
}
int idx = 0;
for ( auto& kv : classmap ) {
unique_labels[idx++] = kv.first;
}
MPI_Op_create( (MPI_User_function *)reduce_unique_labels, 1, &op );
MPI_Allreduce( unique_labels, global_unique_labels, max_size, MPI_INT, op, MPI_COMM_WORLD );
MPI_Op_free( &op );
// update local classmap
std::sort( global_unique_labels, global_unique_labels + max_size );
classmap.clear();
int labeltmp;
idx=0;
for ( int i=0; i<max_size; ++i ) {
labeltmp = global_unique_labels[i];
if ( labeltmp != -1 ) {
classmap.emplace( labeltmp, idx++ );
}
}
// format the local label set into a matrix based on global class map
Mat y = mlu::format_labels( labels, classmap );
numlabels = (LayerSize) classmap.size();
// output total data loading time for each task
MPI_Barrier( MPI_COMM_WORLD );
t2 = MPI_Wtime();
printf( "--- task %d loading time %lf\n", taskid, t2 - t1 );
/* INIT LOCAL CLASSIFIER */
LogisticRegression logistic_layer( n, numlabels, true );
/* OPTIMIZATION */
if ( taskid == MASTER ) {
printf( "\nPerforming Gradient Descent\n" );
}
int update_size; // stores the number of instances read for each update
double grad_mag; // stores the magnitude of the gradient for each update
int delta_size = logistic_layer.get_theta_size();
Vec delta_update = Vec::Zero( delta_size );
int global_update_size;
if ( taskid == MASTER ) {
printf( "iteration : elapsed time : magnitude\n" );
}
for ( int i=0; i<maxiter; ++i ) {
// compute gradient update
t1 = MPI_Wtime();
logistic_layer.compute_gradient( X, y, batch_size, update_size );
delta_data = logistic_layer.get_delta().data();
// sum updates across all partitions
MPI_Allreduce(
delta_data,
delta_update.data(),
delta_size,
MPI_DOUBLE,
MPI_SUM,
MPI_COMM_WORLD
);
logistic_layer.set_delta( delta_update );
// sum the update sizes
MPI_Allreduce( &update_size, &global_update_size, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD );
// normalize + regularize gradient update
logistic_layer.normalize_gradient( global_update_size );
logistic_layer.regularize_gradient( global_update_size );
// update logistic_layer parameters
t2 = MPI_Wtime();
if ( logistic_layer.converged( grad_mag ) ) { break; }
if ( taskid == MASTER ) {
printf( "%d : %lf : %lf\n", i+1, t2 - t1, grad_mag );
}
logistic_layer.update_theta();
}
/* MODEL STORAGE */
if (taskid == MASTER) {
FILE *output;
output = fopen ( output_file.c_str(), "w" );
int idx;
Vec theta = logistic_layer.get_theta();
printf( "\nWriting Model to File: %s\n\n", output_file.c_str() );
fprintf( output, "%lu\n", theta.size() );
for ( idx=0; idx<theta.size()-1; ++idx ) {
fprintf( output, "%lf\t", theta[idx] );
}
fprintf( output, "%lf\n", theta[idx] );
fclose( output );
}
MPI_Finalize();
return 0;
}
