Examples with SoftMax org.nd4j.linalg.api.ops.impl.transforms.SoftMax used on opensource projects

Example 1 with SoftMax

use of org.nd4j.linalg.api.ops.impl.transforms.SoftMax in project deeplearning4j by deeplearning4j.

the class LossFunctionGradientCheck method getFeaturesAndLabels.

public static INDArray[] getFeaturesAndLabels(ILossFunction l, int[] featuresShape, int[] labelsShape, long seed) {
    Nd4j.getRandom().setSeed(seed);
    Random r = new Random(seed);
    INDArray[] ret = new INDArray[2];
    ret[0] = Nd4j.rand(featuresShape);
    switch(l.getClass().getSimpleName()) {
        case "LossBinaryXENT":
            //Want binary vector labels
            ret[1] = Nd4j.rand(labelsShape);
            BooleanIndexing.replaceWhere(ret[1], 0, Conditions.lessThanOrEqual(0.5));
            BooleanIndexing.replaceWhere(ret[1], 1, Conditions.greaterThanOrEqual(0.5));
            break;
        case "LossCosineProximity":
            //Should be real-valued??
            ret[1] = Nd4j.rand(labelsShape).subi(0.5);
            break;
        case "LossKLD":
            //KL divergence: should be a probability distribution for labels??
            ret[1] = Nd4j.rand(labelsShape);
            Nd4j.getExecutioner().exec(new SoftMax(ret[1]), 1);
            break;
        case "LossMCXENT":
        case "LossNegativeLogLikelihood":
            ret[1] = Nd4j.zeros(labelsShape);
            if (labelsShape.length == 2) {
                for (int i = 0; i < labelsShape[0]; i++) {
                    ret[1].putScalar(i, r.nextInt(labelsShape[1]), 1.0);
                }
            } else if (labelsShape.length == 3) {
                for (int i = 0; i < labelsShape[0]; i++) {
                    for (int j = 0; j < labelsShape[2]; j++) {
                        ret[1].putScalar(i, r.nextInt(labelsShape[1]), j, 1.0);
                    }
                }
            } else {
                throw new UnsupportedOperationException();
            }
            break;
        case "LossHinge":
        case "LossSquaredHinge":
            ret[1] = Nd4j.ones(labelsShape);
            if (labelsShape.length == 2) {
                for (int i = 0; i < labelsShape[0]; i++) {
                    ret[1].putScalar(i, r.nextInt(labelsShape[1]), -1.0);
                }
            } else if (labelsShape.length == 3) {
                for (int i = 0; i < labelsShape[0]; i++) {
                    for (int j = 0; j < labelsShape[2]; j++) {
                        ret[1].putScalar(i, r.nextInt(labelsShape[1]), j, -1.0);
                    }
                }
            } else {
                throw new UnsupportedOperationException();
            }
            break;
        case "LossMAPE":
            //requires non-zero values for actual...
            //1 to 2
            ret[1] = Nd4j.rand(labelsShape).addi(1.0);
            break;
        case "LossMAE":
        case "LossMSE":
        case "LossL1":
        case "LossL2":
            ret[1] = Nd4j.rand(labelsShape).muli(2).subi(1);
            break;
        case "LossMSLE":
            //Requires positive labels/activations due to log
            ret[1] = Nd4j.rand(labelsShape);
            break;
        case "LossPoisson":
            //Binary vector labels should be OK here??
            ret[1] = Nd4j.rand(labelsShape);
            BooleanIndexing.replaceWhere(ret[1], 0, Conditions.lessThanOrEqual(0.5));
            BooleanIndexing.replaceWhere(ret[1], 1, Conditions.greaterThanOrEqual(0.5));
            break;
        default:
            throw new IllegalArgumentException("Unknown class: " + l.getClass().getSimpleName());
    }
    return ret;
}

Example 2 with SoftMax

use of org.nd4j.linalg.api.ops.impl.transforms.SoftMax in project deeplearning4j by deeplearning4j.

the class RnnOutputLayer method output.

@Override
public INDArray output(boolean training) {
    //Assume that input is 3d
    if (input.rank() != 3)
        throw new IllegalArgumentException("input must be rank 3");
    INDArray preOutput2d = preOutput2d(training);
    //if(conf.getLayer().getActivationFunction().equals("softmax")) {
    if (conf.getLayer().getActivationFn() instanceof ActivationSoftmax) {
        INDArray out2d = Nd4j.getExecutioner().execAndReturn(new SoftMax(preOutput2d));
        if (maskArray != null) {
            out2d.muliColumnVector(maskArray);
        }
        return TimeSeriesUtils.reshape2dTo3d(out2d, input.size(0));
    }
    if (training)
        applyDropOutIfNecessary(training);
    INDArray origInput = input;
    this.input = TimeSeriesUtils.reshape3dTo2d(input);
    INDArray out = super.activate(true);
    this.input = origInput;
    if (maskArray != null) {
        out.muliColumnVector(maskArray);
    }
    return TimeSeriesUtils.reshape2dTo3d(out, input.size(0));
}