package jsat.classifiers.neuralnetwork.activations;
import jsat.linear.Matrix;
import jsat.linear.Vec;
/**
* This Activation Layer is for Rectified Linear Units. A
* ReLU activation is simply f(x) = max(0, x), and is thus very fast to compute.
*
* See: Nair, V., & Hinton, G. E. (2010). Rectified Linear Units Improve
* Restricted Boltzmann Machines. Proceedings of the 27th International
* Conference on Machine Learning, 807–814.
* @author Edward Raff
*/
public class ReLU implements ActivationLayer
{
private static final long serialVersionUID = -6691240473485759789L;
@Override
public void activate(Vec input, Vec output)
{
for(int i = 0; i < input.length(); i++)
output.set(i, Math.max(0, input.get(i)));
}
@Override
public void activate(Matrix input, Matrix output, boolean rowMajor)
{
for(int i = 0; i < input.rows(); i++)
for(int j = 0; j < input.cols(); j++)
output.set(i, j, Math.max(0, input.get(i, j)));
}
@Override
public void backprop(Vec input, Vec output, Vec delta_partial, Vec errout)
{
for(int i = 0; i < input.length(); i++)
{
double out_i = output.get(i);
if(out_i != 0)
errout.set(i, delta_partial.get(i));
else
errout.set(i, 0.0);
}
}
@Override
public void backprop(Matrix input, Matrix output, Matrix delta_partial, Matrix errout, boolean rowMajor)
{
for (int i = 0; i < input.rows(); i++)
for (int j = 0; j < input.cols(); j++)
if (output.get(i, j) != 0)
errout.set(i, j, delta_partial.get(i, j));
else
errout.set(i, j, 0.0);
}
@Override
public ReLU clone()
{
return new ReLU();
}
}
Ads help maintain this website.