Examples with InnerNode com.simiacryptus.mindseye.network.InnerNode used on opensource projects

Example 1 with InnerNode

use of com.simiacryptus.mindseye.network.InnerNode in project MindsEye by SimiaCryptus.

the class StyleTransfer method getStyleComponents.

/**
 * Gets style components.
 *
 * @param node          the node
 * @param network       the network
 * @param styleParams   the style params
 * @param mean          the mean
 * @param covariance    the covariance
 * @param centeringMode the centering mode
 * @return the style components
 */
@Nonnull
public ArrayList<Tuple2<Double, DAGNode>> getStyleComponents(final DAGNode node, final PipelineNetwork network, final LayerStyleParams styleParams, final Tensor mean, final Tensor covariance, final CenteringMode centeringMode) {
    ArrayList<Tuple2<Double, DAGNode>> styleComponents = new ArrayList<>();
    if (null != styleParams && (styleParams.cov != 0 || styleParams.mean != 0)) {
        double meanRms = mean.rms();
        double meanScale = 0 == meanRms ? 1 : (1.0 / meanRms);
        InnerNode negTarget = network.wrap(new ValueLayer(mean.scale(-1)), new DAGNode[] {});
        InnerNode negAvg = network.wrap(new BandAvgReducerLayer().setAlpha(-1), node);
        if (styleParams.cov != 0) {
            DAGNode recentered;
            switch(centeringMode) {
                case Origin:
                    recentered = node;
                    break;
                case Dynamic:
                    recentered = network.wrap(new GateBiasLayer(), node, negAvg);
                    break;
                case Static:
                    recentered = network.wrap(new GateBiasLayer(), node, negTarget);
                    break;
                default:
                    throw new RuntimeException();
            }
            int[] covDim = covariance.getDimensions();
            assert 0 < covDim[2] : Arrays.toString(covDim);
            int inputBands = mean.getDimensions()[2];
            assert 0 < inputBands : Arrays.toString(mean.getDimensions());
            int outputBands = covDim[2] / inputBands;
            assert 0 < outputBands : Arrays.toString(covDim) + " / " + inputBands;
            double covRms = covariance.rms();
            double covScale = 0 == covRms ? 1 : (1.0 / covRms);
            styleComponents.add(new Tuple2<>(styleParams.cov, network.wrap(new MeanSqLossLayer().setAlpha(covScale), network.wrap(new ValueLayer(covariance), new DAGNode[] {}), network.wrap(ArtistryUtil.wrapTilesAvg(new GramianLayer()), recentered))));
        }
        if (styleParams.mean != 0) {
            styleComponents.add(new Tuple2<>(styleParams.mean, network.wrap(new MeanSqLossLayer().setAlpha(meanScale), negAvg, negTarget)));
        }
    }
    return styleComponents;
}

Example 2 with InnerNode

use of com.simiacryptus.mindseye.network.InnerNode in project MindsEye by SimiaCryptus.

the class ExplodedConvolutionGrid method add.

/**
 * Add dag node.
 *
 * @param input the input
 * @return the dag node
 */
public DAGNode add(@Nonnull DAGNode input) {
    assertAlive();
    DAGNetwork network = input.getNetwork();
    int defaultPaddingX = 0;
    int defaultPaddingY = 0;
    boolean customPaddingX = this.convolutionParams.paddingX != null && convolutionParams.paddingX != defaultPaddingX;
    boolean customPaddingY = this.convolutionParams.paddingY != null && convolutionParams.paddingY != defaultPaddingY;
    final DAGNode paddedInput;
    if (customPaddingX || customPaddingY) {
        int x;
        if (this.convolutionParams.paddingX < -defaultPaddingX) {
            x = this.convolutionParams.paddingX + defaultPaddingX;
        } else if (this.convolutionParams.paddingX > defaultPaddingX) {
            x = this.convolutionParams.paddingX - defaultPaddingX;
        } else {
            x = 0;
        }
        int y;
        if (this.convolutionParams.paddingY < -defaultPaddingY) {
            y = this.convolutionParams.paddingY + defaultPaddingY;
        } else if (this.convolutionParams.paddingY > defaultPaddingY) {
            y = this.convolutionParams.paddingY - defaultPaddingY;
        } else {
            y = 0;
        }
        if (x != 0 || y != 0) {
            paddedInput = network.wrap(new ImgZeroPaddingLayer(x, y).setPrecision(convolutionParams.precision), input);
        } else {
            paddedInput = input;
        }
    } else {
        paddedInput = input;
    }
    InnerNode output;
    if (subLayers.size() == 1) {
        output = (InnerNode) subLayers.get(0).add(paddedInput);
    } else {
        ImgLinearSubnetLayer linearSubnetLayer = new ImgLinearSubnetLayer();
        subLayers.forEach(leg -> {
            PipelineNetwork subnet = new PipelineNetwork();
            leg.add(subnet.getHead());
            linearSubnetLayer.add(leg.fromBand, leg.toBand, subnet);
        });
        boolean isParallel = CudaSettings.INSTANCE.isConv_para_1();
        linearSubnetLayer.setPrecision(convolutionParams.precision).setParallel(isParallel);
        output = network.wrap(linearSubnetLayer, paddedInput).setParallel(isParallel);
    }
    if (customPaddingX || customPaddingY) {
        int x = !customPaddingX ? 0 : (this.convolutionParams.paddingX - defaultPaddingX);
        int y = !customPaddingY ? 0 : (this.convolutionParams.paddingY - defaultPaddingY);
        if (x > 0)
            x = 0;
        if (y > 0)
            y = 0;
        if (x != 0 || y != 0) {
            return network.wrap(new ImgZeroPaddingLayer(x, y).setPrecision(convolutionParams.precision), output);
        }
    }
    return output;
}