com.expleague.ml.methods.seq.nn.LogisticNode Maven / Gradle / Ivy

Go to download

Show more of this group Show more artifacts with this name
Show all versions of jmll

Various ML methods implemented by myself and my students

There is a newer version: 1.4.9

package com.expleague.ml.methods.seq.nn;

import com.expleague.commons.math.vectors.Vec;
import com.expleague.commons.math.vectors.VecTools;
import com.expleague.commons.math.vectors.impl.vectors.ArrayVec;
import com.expleague.commons.math.FuncC1;
import com.expleague.commons.math.vectors.SingleValueVec;
import com.expleague.commons.random.FastRandom;
import com.expleague.ml.func.generic.WSumSigmoid;

public class LogisticNode implements NetworkNode {
  private final Vec w;
  private final FuncC1 func;

  public LogisticNode(int inputDim, FastRandom random) {
    w = new ArrayVec(inputDim + 1);
    for (int i = 0; i < inputDim; i++) {
      w.set(i, random.nextGaussian() / inputDim);
    }
    func = new WSumSigmoid(w);
  }

  @Override
  public Vec params() {
    return w;
  }

  @Override
  public NodeGrad grad(Vec input, Vec nodeOutputGrad) {
    final double outputGrad = nodeOutputGrad.get(0);
    final Vec inputWithOne = VecTools.concat(input, new SingleValueVec(1));

    final Vec inputGrad = func.gradient(inputWithOne);
    VecTools.scale(inputGrad, outputGrad);

    final Vec paramsGrad  = new WSumSigmoid(inputWithOne).gradient(w);
    VecTools.scale(paramsGrad, outputGrad);

    return new NodeGrad(paramsGrad, inputGrad.sub(0, input.dim()));
  }

  @Override
  public Vec value(Vec input) {
    return new SingleValueVec(func.value(VecTools.concat(input, new SingleValueVec(1))));
  }
}