package jsat.datatransform;
import java.util.Iterator;
import jsat.DataSet;
import jsat.classifiers.DataPoint;
import jsat.linear.*;
/**
* Dense sparce transform alters the vectors that store the numerical values.
* Based on a threshold in (0, 1), vectors will be converted from dense to
* sparce, sparce to dense, or left alone.
*
* @author Edward Raff
*/
public class DenseSparceTransform implements DataTransform
{
private static final long serialVersionUID = -1177913691660616290L;
private double factor;
/**
* Creates a new Dense Sparce Transform. The factor gives the maximal
* percentage of values that may be non zero for a vector to be sparce. Any
* vector meeting the requirement will be converted to a sparce vector, and
* others made dense. If the factor is greater than or equal to 1, then all
* vectors will be made sparce. If less than or equal to 0, then all will
* be made dense.
*
* @param factor the fraction of the vectors values that may be non zero to qualify as sparce
*/
public DenseSparceTransform(double factor)
{
this.factor = factor;
}
@Override
public void fit(DataSet data)
{
//no - op, nothing we need to learn
}
@Override
public DataPoint transform(DataPoint dp)
{
Vec orig = dp.getNumericalValues();
final int nnz = orig.nnz();
if (nnz / (double) orig.length() < factor)///make sparse
{
if(orig.isSparse())//already sparse, just return
return dp;
//else, make sparse
SparseVector sv = new SparseVector(orig.length(), nnz);//TODO create a constructor for this
for(int i = 0; i < orig.length(); i++)
if(orig.get(i) != 0)
sv.set(i, orig.get(i));
return new DataPoint(sv, dp.getCategoricalValues(), dp.getCategoricalData(), dp.getWeight());
}
else//make dense
{
if(!orig.isSparse())//already dense, just return
return dp;
DenseVector dv = new DenseVector(orig.length());
Iterator iter = orig.getNonZeroIterator();
while (iter.hasNext())
{
IndexValue indexValue = iter.next();
dv.set(indexValue.getIndex(), indexValue.getValue());
}
return new DataPoint(dv, dp.getCategoricalValues(), dp.getCategoricalData(), dp.getWeight());
}
}
@Override
public DenseSparceTransform clone()
{
return new DenseSparceTransform(factor);
}
}
Ads help maintain this website.