Ejemplo n.º 1
0
  @Override
  public double[][] infer(double[][] features, KObject origin, KInternalDataManager manager) {
    int maxOutput = ((MetaEnum) origin.metaClass().outputs()[0].type()).literals().length;
    KMemoryChunk ks =
        manager.closestChunk(
            origin.universe(),
            origin.now(),
            origin.uuid(),
            origin.metaClass(),
            ((AbstractKObject) origin).previousResolved());
    int dependenciesIndex = origin.metaClass().dependencies().index();
    // check if chunk is empty
    int size = (maxOutput + 1) * (origin.metaClass().inputs().length * NUMOFFIELDS + 1);
    if (ks.getDoubleArraySize(dependenciesIndex, origin.metaClass()) == 0) {
      return null;
    }
    Array1D state =
        new Array1D(size, 0, origin.metaClass().dependencies().index(), ks, origin.metaClass());
    double[][] result = new double[features.length][1];

    for (int j = 0; j < features.length; j++) {
      result[j] = new double[1];
      double maxprob = 0;
      double prob = 0;
      for (int output = 0; output < maxOutput; output++) {
        prob = getProba(features[j], output, state, origin.metaClass().dependencies());
        if (prob > maxprob) {
          maxprob = prob;
          result[j][0] = output;
        }
      }
    }
    return result;
  }
Ejemplo n.º 2
0
  @Override
  public void train(
      double[][] trainingSet,
      double[][] expectedResultSet,
      KObject origin,
      KInternalDataManager manager) {
    int maxOutput = ((MetaEnum) origin.metaClass().outputs()[0].type()).literals().length;
    KMemoryChunk ks =
        manager.preciseChunk(
            origin.universe(),
            origin.now(),
            origin.uuid(),
            origin.metaClass(),
            ((AbstractKObject) origin).previousResolved());
    int dependenciesIndex = origin.metaClass().dependencies().index();
    // Create initial chunk if empty
    int size = (maxOutput + 1) * (origin.metaClass().inputs().length * NUMOFFIELDS + 1);
    if (ks.getDoubleArraySize(dependenciesIndex, origin.metaClass()) == 0) {
      ks.extendDoubleArray(origin.metaClass().dependencies().index(), size, origin.metaClass());
      for (int i = 0; i < size; i++) {
        ks.setDoubleArrayElem(dependenciesIndex, i, 0, origin.metaClass());
      }
    }

    Array1D state =
        new Array1D(size, 0, origin.metaClass().dependencies().index(), ks, origin.metaClass());

    // update the state
    for (int i = 0; i < trainingSet.length; i++) {
      int output = (int) expectedResultSet[i][0];
      for (int j = 0; j < origin.metaClass().inputs().length; j++) {
        // If this is the first datapoint
        if (state.get(getCounter(output, origin.metaClass().dependencies())) == 0) {
          state.set(getIndex(j, output, MIN, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.set(getIndex(j, output, MAX, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.set(getIndex(j, output, SUM, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.set(
              getIndex(j, output, SUMSQUARE, origin.metaClass().dependencies()),
              trainingSet[i][j] * trainingSet[i][j]);

        } else {
          if (trainingSet[i][j]
              < state.get(getIndex(j, output, MIN, origin.metaClass().dependencies()))) {
            state.set(
                getIndex(j, output, MIN, origin.metaClass().dependencies()), trainingSet[i][j]);
          }
          if (trainingSet[i][j]
              > state.get(getIndex(j, output, MAX, origin.metaClass().dependencies()))) {
            state.set(
                getIndex(j, output, MAX, origin.metaClass().dependencies()), trainingSet[i][j]);
          }
          state.add(getIndex(j, output, SUM, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.add(
              getIndex(j, output, SUMSQUARE, origin.metaClass().dependencies()),
              trainingSet[i][j] * trainingSet[i][j]);
        }

        // update global stat
        if (state.get(getCounter(maxOutput, origin.metaClass().dependencies())) == 0) {
          state.set(
              getIndex(j, maxOutput, MIN, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.set(
              getIndex(j, maxOutput, MAX, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.set(
              getIndex(j, maxOutput, SUM, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.set(
              getIndex(j, maxOutput, SUMSQUARE, origin.metaClass().dependencies()),
              trainingSet[i][j] * trainingSet[i][j]);
        } else {
          if (trainingSet[i][j]
              < state.get(getIndex(j, maxOutput, MIN, origin.metaClass().dependencies()))) {
            state.set(
                getIndex(j, maxOutput, MIN, origin.metaClass().dependencies()), trainingSet[i][j]);
          }
          if (trainingSet[i][j]
              > state.get(getIndex(j, maxOutput, MAX, origin.metaClass().dependencies()))) {
            state.set(
                getIndex(j, maxOutput, MAX, origin.metaClass().dependencies()), trainingSet[i][j]);
          }
          state.add(
              getIndex(j, maxOutput, SUM, origin.metaClass().dependencies()), trainingSet[i][j]);
          state.add(
              getIndex(j, maxOutput, SUMSQUARE, origin.metaClass().dependencies()),
              trainingSet[i][j] * trainingSet[i][j]);
        }
      }

      // Update Global counters
      state.add(getCounter(output, origin.metaClass().dependencies()), 1);
      state.add(getCounter(maxOutput, origin.metaClass().dependencies()), 1);
    }
  }