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

Example 16 with PipelineNetwork

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

the class ConvolutionLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    final Tensor kernel = getKernel();
    kernel.addRef();
    assert kernel.isValid();
    assert 1 == inObj.length;
    assert 3 == inObj[0].getData().getDimensions().length;
    assert inputBands == inObj[0].getData().getDimensions()[2] : Arrays.toString(inObj[0].getData().getDimensions()) + "[2] != " + inputBands;
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    @Nonnull ExplodedConvolutionGrid grid = getExplodedNetwork();
    @Nonnull PipelineNetwork network = grid.getNetwork();
    if (isFrozen()) {
        network.freeze();
    }
    final Result result = network.evalAndFree(inObj);
    network.freeRef();
    final TensorList resultData = result.getData();
    assert inObj[0].getData().length() == resultData.length();
    assert 3 == resultData.getDimensions().length;
    assert outputBands == resultData.getDimensions()[2];
    ConvolutionLayer.this.addRef();
    return new Result(resultData, (@Nonnull final DeltaSet<Layer> deltaSet, @Nonnull final TensorList delta) -> {
        result.accumulate(deltaSet, delta);
        if (!isFrozen()) {
            Tensor read = grid.read(deltaSet, true);
            deltaSet.get(ConvolutionLayer.this, kernel.getData()).addInPlace(read.getData()).freeRef();
            read.freeRef();
        }
    }) {

        @Override
        public void accumulate(final DeltaSet<Layer> buffer, final TensorList delta) {
            getAccumulator().accept(buffer, delta);
        }

        @Override
        protected void _free() {
            grid.freeRef();
            result.freeRef();
            kernel.freeRef();
            ConvolutionLayer.this.freeRef();
        }

        @Override
        public boolean isAlive() {
            return result.isAlive();
        }
    };
}

Example 17 with PipelineNetwork

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

the class ArtistryUtil method wrapTilesAvg.

/**
 * Wrap tiles avg layer.
 *
 * @param subnet     the subnet
 * @param borderX1   the border x 1
 * @param borderY1   the border y 1
 * @param borderX2   the border x 2
 * @param borderY2   the border y 2
 * @param tileWidth  the tile width
 * @param tileHeight the tile height
 * @return the layer
 */
protected static Layer wrapTilesAvg(final Layer subnet, final int borderX1, final int borderY1, final int borderX2, final int borderY2, final int tileWidth, final int tileHeight) {
    PipelineNetwork network1 = new PipelineNetwork(1);
    if (borderX1 != 0 || borderY1 != 0)
        network1.wrap(new com.simiacryptus.mindseye.layers.cudnn.ImgZeroPaddingLayer(borderX1, borderY1));
    network1.add(subnet);
    if (borderX2 != 0 || borderY2 != 0)
        network1.wrap(new com.simiacryptus.mindseye.layers.cudnn.ImgZeroPaddingLayer(-borderX2, -borderY2));
    PipelineNetwork network = new PipelineNetwork(1);
    network.wrap(new com.simiacryptus.mindseye.layers.cudnn.ImgTileSubnetLayer(network1, tileWidth, tileHeight, tileWidth - 2 * borderX1, tileHeight - 2 * borderY1));
    network.wrap(new BandAvgReducerLayer());
    return network;
}

Example 18 with PipelineNetwork

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

the class AutoencodingProblem method run.

@Nonnull
@Override
public AutoencodingProblem run(@Nonnull final NotebookOutput log) {
    @Nonnull final DAGNetwork fwdNetwork = fwdFactory.imageToVector(log, features);
    @Nonnull final DAGNetwork revNetwork = revFactory.vectorToImage(log, features);
    @Nonnull final PipelineNetwork echoNetwork = new PipelineNetwork(1);
    echoNetwork.add(fwdNetwork);
    echoNetwork.add(revNetwork);
    @Nonnull final PipelineNetwork supervisedNetwork = new PipelineNetwork(1);
    supervisedNetwork.add(fwdNetwork);
    @Nonnull final DropoutNoiseLayer dropoutNoiseLayer = new DropoutNoiseLayer().setValue(dropout);
    supervisedNetwork.add(dropoutNoiseLayer);
    supervisedNetwork.add(revNetwork);
    supervisedNetwork.add(new MeanSqLossLayer(), supervisedNetwork.getHead(), supervisedNetwork.getInput(0));
    log.h3("Network Diagrams");
    log.code(() -> {
        return Graphviz.fromGraph(TestUtil.toGraph(fwdNetwork)).height(400).width(600).render(Format.PNG).toImage();
    });
    log.code(() -> {
        return Graphviz.fromGraph(TestUtil.toGraph(revNetwork)).height(400).width(600).render(Format.PNG).toImage();
    });
    log.code(() -> {
        return Graphviz.fromGraph(TestUtil.toGraph(supervisedNetwork)).height(400).width(600).render(Format.PNG).toImage();
    });
    @Nonnull final TrainingMonitor monitor = new TrainingMonitor() {

        @Nonnull
        TrainingMonitor inner = TestUtil.getMonitor(history);

        @Override
        public void log(final String msg) {
            inner.log(msg);
        }

        @Override
        public void onStepComplete(final Step currentPoint) {
            dropoutNoiseLayer.shuffle(StochasticComponent.random.get().nextLong());
            inner.onStepComplete(currentPoint);
        }
    };
    final Tensor[][] trainingData = getTrainingData(log);
    // MonitoredObject monitoringRoot = new MonitoredObject();
    // TestUtil.addMonitoring(supervisedNetwork, monitoringRoot);
    log.h3("Training");
    TestUtil.instrumentPerformance(supervisedNetwork);
    @Nonnull final ValidatingTrainer trainer = optimizer.train(log, new SampledArrayTrainable(trainingData, supervisedNetwork, trainingData.length / 2, batchSize), new ArrayTrainable(trainingData, supervisedNetwork, batchSize), monitor);
    log.code(() -> {
        trainer.setTimeout(timeoutMinutes, TimeUnit.MINUTES).setMaxIterations(10000).run();
    });
    if (!history.isEmpty()) {
        log.code(() -> {
            return TestUtil.plot(history);
        });
        log.code(() -> {
            return TestUtil.plotTime(history);
        });
    }
    TestUtil.extractPerformance(log, supervisedNetwork);
    {
        @Nonnull final String modelName = "encoder_model" + AutoencodingProblem.modelNo++ + ".json";
        log.p("Saved model as " + log.file(fwdNetwork.getJson().toString(), modelName, modelName));
    }
    @Nonnull final String modelName = "decoder_model" + AutoencodingProblem.modelNo++ + ".json";
    log.p("Saved model as " + log.file(revNetwork.getJson().toString(), modelName, modelName));
    // log.h3("Metrics");
    // log.code(() -> {
    // return TestUtil.toFormattedJson(monitoringRoot.getMetrics());
    // });
    log.h3("Validation");
    log.p("Here are some re-encoded examples:");
    log.code(() -> {
        @Nonnull final TableOutput table = new TableOutput();
        data.validationData().map(labeledObject -> {
            return toRow(log, labeledObject, echoNetwork.eval(labeledObject.data).getData().get(0).getData());
        }).filter(x -> null != x).limit(10).forEach(table::putRow);
        return table;
    });
    log.p("Some rendered unit vectors:");
    for (int featureNumber = 0; featureNumber < features; featureNumber++) {
        @Nonnull final Tensor input = new Tensor(features).set(featureNumber, 1);
        @Nullable final Tensor tensor = revNetwork.eval(input).getData().get(0);
        log.out(log.image(tensor.toImage(), ""));
    }
    return this;
}

Example 19 with PipelineNetwork

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

the class EncodingProblem method run.

@Nonnull
@Override
public EncodingProblem run(@Nonnull final NotebookOutput log) {
    @Nonnull final TrainingMonitor monitor = TestUtil.getMonitor(history);
    Tensor[][] trainingData;
    try {
        trainingData = data.trainingData().map(labeledObject -> {
            return new Tensor[] { new Tensor(features).set(this::random), labeledObject.data };
        }).toArray(i -> new Tensor[i][]);
    } catch (@Nonnull final IOException e) {
        throw new RuntimeException(e);
    }
    @Nonnull final DAGNetwork imageNetwork = revFactory.vectorToImage(log, features);
    log.h3("Network Diagram");
    log.code(() -> {
        return Graphviz.fromGraph(TestUtil.toGraph(imageNetwork)).height(400).width(600).render(Format.PNG).toImage();
    });
    @Nonnull final PipelineNetwork trainingNetwork = new PipelineNetwork(2);
    @Nullable final DAGNode image = trainingNetwork.add(imageNetwork, trainingNetwork.getInput(0));
    @Nullable final DAGNode softmax = trainingNetwork.add(new SoftmaxActivationLayer(), trainingNetwork.getInput(0));
    trainingNetwork.add(new SumInputsLayer(), trainingNetwork.add(new EntropyLossLayer(), softmax, softmax), trainingNetwork.add(new NthPowerActivationLayer().setPower(1.0 / 2.0), trainingNetwork.add(new MeanSqLossLayer(), image, trainingNetwork.getInput(1))));
    log.h3("Training");
    log.p("We start by training apply a very small population to improve initial convergence performance:");
    TestUtil.instrumentPerformance(trainingNetwork);
    @Nonnull final Tensor[][] primingData = Arrays.copyOfRange(trainingData, 0, 1000);
    @Nonnull final ValidatingTrainer preTrainer = optimizer.train(log, (SampledTrainable) new SampledArrayTrainable(primingData, trainingNetwork, trainingSize, batchSize).setMinSamples(trainingSize).setMask(true, false), new ArrayTrainable(primingData, trainingNetwork, batchSize), monitor);
    log.code(() -> {
        preTrainer.setTimeout(timeoutMinutes / 2, TimeUnit.MINUTES).setMaxIterations(batchSize).run();
    });
    TestUtil.extractPerformance(log, trainingNetwork);
    log.p("Then our main training phase:");
    TestUtil.instrumentPerformance(trainingNetwork);
    @Nonnull final ValidatingTrainer mainTrainer = optimizer.train(log, (SampledTrainable) new SampledArrayTrainable(trainingData, trainingNetwork, trainingSize, batchSize).setMinSamples(trainingSize).setMask(true, false), new ArrayTrainable(trainingData, trainingNetwork, batchSize), monitor);
    log.code(() -> {
        mainTrainer.setTimeout(timeoutMinutes, TimeUnit.MINUTES).setMaxIterations(batchSize).run();
    });
    TestUtil.extractPerformance(log, trainingNetwork);
    if (!history.isEmpty()) {
        log.code(() -> {
            return TestUtil.plot(history);
        });
        log.code(() -> {
            return TestUtil.plotTime(history);
        });
    }
    try {
        @Nonnull String filename = log.getName().toString() + EncodingProblem.modelNo++ + "_plot.png";
        ImageIO.write(Util.toImage(TestUtil.plot(history)), "png", log.file(filename));
        log.appendFrontMatterProperty("result_plot", filename, ";");
    } catch (IOException e) {
        throw new RuntimeException(e);
    }
    // log.file()
    @Nonnull final String modelName = "encoding_model_" + EncodingProblem.modelNo++ + ".json";
    log.appendFrontMatterProperty("result_model", modelName, ";");
    log.p("Saved model as " + log.file(trainingNetwork.getJson().toString(), modelName, modelName));
    log.h3("Results");
    @Nonnull final PipelineNetwork testNetwork = new PipelineNetwork(2);
    testNetwork.add(imageNetwork, testNetwork.getInput(0));
    log.code(() -> {
        @Nonnull final TableOutput table = new TableOutput();
        Arrays.stream(trainingData).map(tensorArray -> {
            @Nullable final Tensor predictionSignal = testNetwork.eval(tensorArray).getData().get(0);
            @Nonnull final LinkedHashMap<CharSequence, Object> row = new LinkedHashMap<>();
            row.put("Source", log.image(tensorArray[1].toImage(), ""));
            row.put("Echo", log.image(predictionSignal.toImage(), ""));
            return row;
        }).filter(x -> null != x).limit(10).forEach(table::putRow);
        return table;
    });
    log.p("Learned Model Statistics:");
    log.code(() -> {
        @Nonnull final ScalarStatistics scalarStatistics = new ScalarStatistics();
        trainingNetwork.state().stream().flatMapToDouble(x -> Arrays.stream(x)).forEach(v -> scalarStatistics.add(v));
        return scalarStatistics.getMetrics();
    });
    log.p("Learned Representation Statistics:");
    log.code(() -> {
        @Nonnull final ScalarStatistics scalarStatistics = new ScalarStatistics();
        Arrays.stream(trainingData).flatMapToDouble(row -> Arrays.stream(row[0].getData())).forEach(v -> scalarStatistics.add(v));
        return scalarStatistics.getMetrics();
    });
    log.p("Some rendered unit vectors:");
    for (int featureNumber = 0; featureNumber < features; featureNumber++) {
        @Nonnull final Tensor input = new Tensor(features).set(featureNumber, 1);
        @Nullable final Tensor tensor = imageNetwork.eval(input).getData().get(0);
        TestUtil.renderToImages(tensor, true).forEach(img -> {
            log.out(log.image(img, ""));
        });
    }
    return this;
}

Example 20 with PipelineNetwork

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

the class AutoencoderNetwork method train.

/**
 * Train autoencoder network . training parameters.
 *
 * @return the autoencoder network . training parameters
 */
@Nonnull
public AutoencoderNetwork.TrainingParameters train() {
    return new AutoencoderNetwork.TrainingParameters() {

        @Nonnull
        @Override
        public SimpleLossNetwork getTrainingNetwork() {
            @Nonnull final PipelineNetwork student = new PipelineNetwork();
            student.add(encoder);
            student.add(decoder);
            return new SimpleLossNetwork(student, new MeanSqLossLayer());
        }

        @Nonnull
        @Override
        protected TrainingMonitor wrap(@Nonnull final TrainingMonitor monitor) {
            return new TrainingMonitor() {

                @Override
                public void log(final String msg) {
                    monitor.log(msg);
                }

                @Override
                public void onStepComplete(final Step currentPoint) {
                    inputNoise.shuffle();
                    encodedNoise.shuffle(StochasticComponent.random.get().nextLong());
                    monitor.onStepComplete(currentPoint);
                }
            };
        }
    };
}