com.expleague.ml.methods.seq.param.WeightSquareParametrization 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.param;

import com.expleague.commons.math.MathTools;
import com.expleague.commons.math.vectors.Mx;
import com.expleague.commons.math.vectors.Vec;
import com.expleague.commons.math.vectors.VecTools;
import com.expleague.commons.math.vectors.impl.mx.VecBasedMx;

public class WeightSquareParametrization implements WeightParametrization{
  private final BettaParametrization bettaParametrization;

  public WeightSquareParametrization(BettaParametrization bettaParametrization) {
    this.bettaParametrization = bettaParametrization;
  }

  @Override
  public Mx getMx(Vec params, int c, int stateCount) {
    Mx betta = bettaParametrization.getBettaMx(params, c, stateCount);
    return getMx(betta, c, stateCount);
  }

  private Mx getMx(Mx betta, int c, int stateCount) {
    Mx result = new VecBasedMx(stateCount, stateCount);

    for (int f = 0; f < stateCount; f++) {
      double sum = 0;
      for (int j = 0; j < stateCount; j++) {
        final double e = MathTools.sqr(betta.get(f, j));
        sum += e;
        result.set(j, f, e);
      }

      for (int t = 0; t < stateCount; t++) {
        result.set(t, f, result.get(t, f) / sum);
      }
    }

    return result;
  }

  @Override
  public void gradientTo(Vec params, Vec out, Vec dOut, Vec dOutNew, Mx dW, int c, int stateCount) {
    VecTools.fill(dW, 0);

    final Mx betta = bettaParametrization.getBettaMx(params, c, stateCount);
    final Mx W = getMx(betta, c, stateCount);

    for (int f = 0; f < stateCount; f++) {
      final double prevS_f = out.get(f);

      double sum = 0;
      for (int i = 0; i < stateCount; i++) {
        sum += MathTools.sqr(betta.get(f, i));
      }

      double prevdS_f = 0;
      for (int t = 0; t < stateCount; t++) {
        { // dT/dS_prev_f
          prevdS_f += W.get(t, f) * dOut.get(t);
        }

        { // dT/dM
          final double grad = prevS_f * dOut.get(t);
          final double betta_ft = betta.get(f, t);
          final double betta_ft_2 = MathTools.sqr(betta_ft);
          final double multiply = 2 * grad  / sum / sum;

          for (int j = 0; j < stateCount; j++) {
            final double betta_fj = betta.get(f, j);
            if (j == t) {
              dW.adjust(f, j, multiply * betta_fj * (sum - MathTools.sqr(betta_fj)));
            }
            else {
              dW.adjust(f, j, -multiply * betta_ft_2 * betta_fj);
            }
          }
        }
      }
      dOutNew.set(f, prevdS_f);
    }
  }
}