Examples with ReferenceCounting com.simiacryptus.mindseye.lang.ReferenceCounting used on opensource projects

Example 11 with ReferenceCounting

use of com.simiacryptus.mindseye.lang.ReferenceCounting in project MindsEye by SimiaCryptus.

the class AvgReducerLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    int length = inputData.length();
    CudaTensorList result = CudaSystem.run(gpu -> {
        CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        inputData.freeRef();
        CudaMemory inputMemory = inputTensor.getMemory(gpu);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, 1, 1, 1);
        long size = (long) precision.size * outputDescriptor.nStride * length;
        @Nonnull final CudaMemory outputMemory = gpu.allocate(size, MemoryType.Managed, true);
        CudaResource<cudnnReduceTensorDescriptor> reduceTensorDescriptor = gpu.cudnnCreateReduceTensorDescriptor(cudnnReduceTensorOp.CUDNN_REDUCE_TENSOR_AVG, precision.code, cudnnNanPropagation.CUDNN_NOT_PROPAGATE_NAN, cudnnReduceTensorIndices.CUDNN_REDUCE_TENSOR_NO_INDICES, cudnnIndicesType.CUDNN_32BIT_INDICES);
        @Nonnull final CudaMemory workspacePtr = gpu.allocate(inputMemory.size, MemoryType.Device, true);
        @Nonnull final CudaMemory indexPtr = gpu.allocate(12 * length, MemoryType.Device, false);
        // outputPtr.synchronize();
        gpu.cudnnReduceTensor(reduceTensorDescriptor.getPtr(), indexPtr.getPtr(), indexPtr.size, workspacePtr.getPtr(), workspacePtr.size, precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputMemory.getPtr());
        outputMemory.dirty();
        inputMemory.dirty();
        Stream.of(inputTensor, inputMemory, reduceTensorDescriptor, workspacePtr, indexPtr).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(CudaTensor.wrap(outputMemory, outputDescriptor, precision), length, new int[] { 1, 1, 1 }, precision);
    });
    return new Result(result, (DeltaSet<Layer> ctx, TensorList delta) -> {
        // Not supported by CuDNN?
        // CudaTensorList passback = CudaSystem.run(gpu -> {
        // CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
        // CudaMemory deltaMemory = deltaTensor.getMemory(gpu);
        // 
        // @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor1 = gpu.newTensorDescriptor(
        // precision, length, inputSize[2], inputSize[1], inputSize[0]
        // );
        // @Nonnull final CudaMemory passbackPtr1 = gpu.allocate((long) precision.size * passbackDescriptor1.nStride * length, MemoryType.Device, false);
        // gpu.cudnnAddTensor(precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaMemory.getPtr(),
        // precision.getPointer(1.0), passbackDescriptor1.getPtr(), passbackPtr1.getPtr());
        // passbackPtr1.dirty();
        // 
        // Stream.of(deltaTensor, deltaMemory, passbackDescriptor1, passbackPtr1).forEach(ReferenceCounting::freeRef);
        // return CudaTensorList.wrap(CudaTensor.wrap(passbackPtr1, passbackDescriptor1, precision), length, inputSize, precision);
        // });
        TensorList passback = TensorArray.wrap(IntStream.range(0, length).mapToObj(i -> {
            Tensor tensor = delta.get(i);
            Tensor tensor1 = new Tensor(inputSize).setAll((double) tensor.get(0) / Tensor.length(inputSize));
            tensor.freeRef();
            return tensor1;
        }).toArray(i -> new Tensor[i]));
        input.accumulate(ctx, passback);
    }) {

        @Override
        protected void _free() {
            super._free();
            input.freeRef();
        }
    };
}

Example 12 with ReferenceCounting

use of com.simiacryptus.mindseye.lang.ReferenceCounting in project MindsEye by SimiaCryptus.

the class CudaLayerTester method testNonstandardBoundsBackprop.

/**
 * Test nonstandard bounds backprop tolerance statistics.
 *
 * @param log            the log
 * @param layer          the layer
 * @param inputPrototype the input prototype
 * @return the tolerance statistics
 */
@Nonnull
public ToleranceStatistics testNonstandardBoundsBackprop(final NotebookOutput log, @Nullable final Layer layer, @Nonnull final Tensor[] inputPrototype) {
    log.h2("Irregular Backprop");
    log.p("This layer should accept non-dense tensors as delta input.");
    return log.code(() -> {
        Tensor[] randomized = Arrays.stream(inputPrototype).map(x -> x.map(v -> getRandom())).toArray(i -> new Tensor[i]);
        logger.info("Input: " + Arrays.stream(randomized).map(Tensor::prettyPrint).collect(Collectors.toList()));
        Precision precision = Precision.Double;
        TensorList[] controlInput = Arrays.stream(randomized).map(original -> {
            return TensorArray.wrap(original);
        }).toArray(i -> new TensorList[i]);
        @Nonnull final SimpleResult testResult = CudaSystem.run(gpu -> {
            TensorList[] copy = copy(controlInput);
            SimpleResult result = new SimpleGpuEval(layer, gpu, copy) {

                @Nonnull
                @Override
                public TensorList getFeedback(@Nonnull final TensorList original) {
                    Tensor originalTensor = original.get(0).mapAndFree(x -> 1);
                    CudaTensorList cudaTensorList = buildIrregularCudaTensor(gpu, precision, originalTensor);
                    originalTensor.freeRef();
                    return cudaTensorList;
                }
            }.call();
            Arrays.stream(copy).forEach(ReferenceCounting::freeRef);
            return result;
        });
        @Nonnull final SimpleResult controlResult = CudaSystem.run(gpu -> {
            TensorList[] copy = copy(controlInput);
            SimpleResult result = SimpleGpuEval.run(layer, gpu, copy);
            Arrays.stream(copy).forEach(ReferenceCounting::freeRef);
            return result;
        }, 1);
        try {
            ToleranceStatistics compareOutput = compareOutput(controlResult, testResult);
            ToleranceStatistics compareDerivatives = compareDerivatives(controlResult, testResult);
            return compareDerivatives.combine(compareOutput);
        } finally {
            Arrays.stream(controlInput).forEach(ReferenceCounting::freeRef);
            controlResult.freeRef();
            testResult.freeRef();
        }
    });
}

Example 13 with ReferenceCounting

use of com.simiacryptus.mindseye.lang.ReferenceCounting in project MindsEye by SimiaCryptus.

the class CudaLayerTester method testInterGpu.

/**
 * Test inter gpu tolerance statistics.
 *
 * @param log            the log
 * @param reference      the reference
 * @param inputPrototype the input prototype
 * @return the tolerance statistics
 */
@Nonnull
public ToleranceStatistics testInterGpu(final NotebookOutput log, @Nullable final Layer reference, @Nonnull final Tensor[] inputPrototype) {
    log.h2("Multi-GPU Compatibility");
    log.p("This layer should be able to apply using a GPU context other than the one used to create the inputs.");
    return log.code(() -> {
        final TensorList[] heapInput = Arrays.stream(inputPrototype).map(t -> TensorArray.wrap(IntStream.range(0, getBatchSize()).mapToObj(i -> t.map(v -> getRandom())).toArray(i -> new Tensor[i]))).toArray(i -> new TensorList[i]);
        logger.info("Input: " + Arrays.stream(heapInput).flatMap(x -> x.stream()).map(tensor -> {
            String prettyPrint = tensor.prettyPrint();
            tensor.freeRef();
            return prettyPrint;
        }).collect(Collectors.toList()));
        TensorList[] gpuInput = CudaSystem.run(gpu -> {
            return Arrays.stream(heapInput).map(original -> {
                return CudaTensorList.wrap(gpu.getTensor(original, Precision.Double, MemoryType.Managed, false), original.length(), original.getDimensions(), Precision.Double);
            }).toArray(i -> new TensorList[i]);
        }, 0);
        @Nonnull final SimpleResult fromHeap = CudaSystem.run(gpu -> {
            return SimpleGpuEval.run(reference, gpu, heapInput);
        }, 1);
        @Nonnull final SimpleResult fromGPU = CudaSystem.run(gpu -> {
            return SimpleGpuEval.run(reference, gpu, gpuInput);
        }, 1);
        try {
            ToleranceStatistics compareOutput = compareOutput(fromHeap, fromGPU);
            ToleranceStatistics compareDerivatives = compareDerivatives(fromHeap, fromGPU);
            return compareDerivatives.combine(compareOutput);
        } finally {
            Arrays.stream(gpuInput).forEach(ReferenceCounting::freeRef);
            Arrays.stream(heapInput).forEach(x -> x.freeRef());
            fromGPU.freeRef();
            fromHeap.freeRef();
        }
    });
}

Example 14 with ReferenceCounting

use of com.simiacryptus.mindseye.lang.ReferenceCounting in project MindsEye by SimiaCryptus.

the class StandardLayerTests method getInvocations.

/**
 * Gets invocations.
 *
 * @param smallLayer the small layer
 * @param smallDims  the small dims
 * @return the invocations
 */
@Nonnull
public Collection<Invocation> getInvocations(@Nonnull Layer smallLayer, @Nonnull int[][] smallDims) {
    @Nonnull DAGNetwork smallCopy = (DAGNetwork) smallLayer.copy();
    @Nonnull HashSet<Invocation> invocations = new HashSet<>();
    smallCopy.visitNodes(node -> {
        @Nullable Layer inner = node.getLayer();
        inner.addRef();
        @Nullable Layer wrapper = new LayerBase() {

            @Nullable
            @Override
            public Result eval(@Nonnull Result... array) {
                if (null == inner)
                    return null;
                @Nullable Result result = inner.eval(array);
                invocations.add(new Invocation(inner, Arrays.stream(array).map(x -> x.getData().getDimensions()).toArray(i -> new int[i][])));
                return result;
            }

            @Override
            public JsonObject getJson(Map<CharSequence, byte[]> resources, DataSerializer dataSerializer) {
                return inner.getJson(resources, dataSerializer);
            }

            @Nullable
            @Override
            public List<double[]> state() {
                return inner.state();
            }

            @Override
            protected void _free() {
                inner.freeRef();
            }
        };
        node.setLayer(wrapper);
        wrapper.freeRef();
    });
    Tensor[] input = Arrays.stream(smallDims).map(i -> new Tensor(i)).toArray(i -> new Tensor[i]);
    try {
        Result eval = smallCopy.eval(input);
        eval.freeRef();
        eval.getData().freeRef();
        return invocations;
    } finally {
        Arrays.stream(input).forEach(ReferenceCounting::freeRef);
        smallCopy.freeRef();
    }
}

Example 15 with ReferenceCounting

use of com.simiacryptus.mindseye.lang.ReferenceCounting in project MindsEye by SimiaCryptus.

the class TrainingTester method testCompleteLearning.

/**
 * Test complete learning apply result.
 *
 * @param log            the log
 * @param component      the component
 * @param random         the random
 * @param inputPrototype the input prototype
 * @return the apply result
 */
@Nonnull
public TestResult testCompleteLearning(@Nonnull final NotebookOutput log, @Nonnull final Layer component, final Random random, @Nonnull final Tensor[] inputPrototype) {
    @Nonnull final Layer network_target = shuffle(random, component.copy()).freeze();
    final Tensor[][] input_target = shuffleCopy(random, inputPrototype);
    log.p("In this apply, attempt to train a network to emulate a randomized network given an example input/output. The target state is:");
    log.code(() -> {
        return network_target.state().stream().map(Arrays::toString).reduce((a, b) -> a + "\n" + b).orElse("");
    });
    log.p("We simultaneously regress this target input:");
    log.code(() -> {
        return Arrays.stream(input_target).flatMap(x -> Arrays.stream(x)).map(x -> x.prettyPrint()).reduce((a, b) -> a + "\n" + b).orElse("");
    });
    log.p("Which produces the following output:");
    Result[] inputs = ConstantResult.batchResultArray(input_target);
    Result eval = network_target.eval(inputs);
    network_target.freeRef();
    Arrays.stream(inputs).forEach(ReferenceCounting::freeRef);
    TensorList result = eval.getData();
    eval.freeRef();
    final Tensor[] output_target = result.stream().toArray(i -> new Tensor[i]);
    log.code(() -> {
        return Stream.of(output_target).map(x -> x.prettyPrint()).reduce((a, b) -> a + "\n" + b).orElse("");
    });
    // if (output_target.length != inputPrototype.length) return null;
    return trainAll("Integrated Convergence", log, append(shuffleCopy(random, inputPrototype), output_target), shuffle(random, component.copy()), buildMask(inputPrototype.length));
}