Method transform

JSAT/src/jsat/classifiers/neuralnetwork/RBFNet.java:224–271

(DataPoint dp)

    }

    @Override
    public DataPoint transform(DataPoint dp)
    {
        final Vec x = dp.getNumericalValues();
        final List<Double> qi = dm.getQueryInfo(x);
        Vec sv = new SparseVector(numCentroids);
        double sum = 0;
        /*
         * Keep track of the highest activation in case none of the neurons have
         * a numericaly stable activation value. if this occurs we will do our 
         * best by simply setting the one largest activation
         */
        double maxActivation = Double.NEGATIVE_INFINITY;
        int highestNeuron = -1;
                
        for(int i = 0; i < centroids.size(); i++)
        {
            double dist = dm.dist(i, x, qi, centroids, centroidDistCache);
            double sig = bandwidths[i];
            double activation = Math.exp(-(dist*dist)/(sig*sig*2));
            
            if(activation > maxActivation)
            {
                maxActivation = activation;
                highestNeuron = i;
            }
            
            if(activation > 1e-16)
            {
                sv.set(i, activation);
                sum += activation;
            }
        }
        
        if(sv.nnz() == 0)//no activations
        {
            sv.set(highestNeuron, maxActivation);
            sum = maxActivation;
        }
            
        
        if(normalize && sum != 0.0)//-0.0 not an issue with rbf kernel
            sv.mutableDivide(sum);
        if(sv.nnz() > sv.length()/2)//at this point we would be using more memory than needed. Just switch to dense
            sv = new DenseVector(sv);
        
        return new DataPoint(sv, dp.getCategoricalValues(), dp.getCategoricalData(), dp.getWeight());
    }
    /**
     * The first phase of learning a  RBF Neural Network is to determine the 
     * neuron locations. This enum controls which method is used.