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

Example 6 with DeltaSet

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

the class BandAvgReducerLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    final Result input = inObj[0];
    TensorList inputData = input.getData();
    @Nonnull final int[] inputSize = inputData.getDimensions();
    int length = inputData.length();
    final int bands = inputSize[2];
    CudaTensorList result = CudaSystem.run(gpu -> {
        CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, bands, 1, 1);
        long size = (long) precision.size * outputDescriptor.nStride * length;
        @Nonnull final CudaMemory outputPtr = 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);
        CudaMemory inputMemory = inputTensor.getMemory(gpu);
        @Nonnull final CudaMemory workspacePtr = gpu.allocate(inputMemory.size, MemoryType.Device, true);
        @Nonnull final CudaMemory indexPtr = gpu.allocate(12 * length, MemoryType.Device, false);
        gpu.cudnnReduceTensor(reduceTensorDescriptor.getPtr(), indexPtr.getPtr(), indexPtr.size, workspacePtr.getPtr(), workspacePtr.size, precision.getPointer(alpha), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr());
        outputPtr.dirty();
        inputMemory.dirty();
        Stream.of(inputMemory, inputTensor, reduceTensorDescriptor, workspacePtr, indexPtr, inputData).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, new int[] { 1, 1, bands }, precision);
    });
    int pixels = inputSize[0] * inputSize[1];
    return new Result(result, (DeltaSet<Layer> ctx, TensorList delta) -> {
        TensorList passback;
        passback = TensorArray.wrap(delta.stream().map(x -> {
            Tensor tensor = new Tensor(inputSize[0], inputSize[1], inputSize[2]).setByCoord(c -> x.get(c.getCoords()[2]) * alpha / pixels);
            x.freeRef();
            return tensor;
        }).toArray(i -> new Tensor[i]));
        // passback = CudaSystem.run(gpu -> {
        // CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
        // @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision,
        // length, inputSize[2], inputSize[1], inputSize[0]);
        // @Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Device, true);
        // CudaMemory deltaMemory = deltaTensor.getMemory(gpu);
        // @Nonnull final CudaDevice.CudaTensorDescriptor inputDescriptor = gpu.newTensorDescriptor(precision,
        // 1, 1, inputSize[1], inputSize[0]);
        // for(int batch=0;batch<length;batch++){
        // Tensor tensor = delta.get(batch);
        // for(int band=0;band<bands;band++){
        // int i = batch * bands + band;
        // CudaMemory img = outputPtr.withByteOffset(precision.size * i * outputDescriptor.cStride);
        // CudaMemory val = deltaMemory.withByteOffset(precision.size * i);
        // gpu.cudnnSetTensor(inputDescriptor.getPtr(), img.getPtr(), precision.getPointer(tensor.get(band) / outputDescriptor.cStride));
        // img.freeRef();
        // val.freeRef();
        // outputPtr.dirty().synchronize();
        // }
        // }
        // Stream.of(deltaMemory, deltaTensor, inputDescriptor).forEach(ReferenceCounting::freeRef);
        // return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, inputSize, precision);
        // });
        input.accumulate(ctx, passback);
    }) {

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

Example 7 with DeltaSet

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

the class BinarySumLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (inObj.length == 1) {
        if (rightFactor != 1)
            throw new IllegalStateException();
        if (leftFactor != 1)
            throw new IllegalStateException();
        return inObj[0];
    }
    if (inObj.length > 2) {
        if (rightFactor != 1)
            throw new IllegalStateException();
        if (leftFactor != 1)
            throw new IllegalStateException();
        return Arrays.stream(inObj).reduce((a, b) -> evalAndFree(a, b)).get();
    }
    assert (inObj.length == 2);
    final TensorList leftData = inObj[0].getData();
    final TensorList rightData = inObj[1].getData();
    int[] leftDimensions = leftData.getDimensions();
    if (3 < leftDimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(leftDimensions));
    }
    @Nonnull final int[] dimensions = { leftDimensions.length < 1 ? 0 : leftDimensions[0], leftDimensions.length < 2 ? 1 : leftDimensions[1], leftDimensions.length < 3 ? 1 : leftDimensions[2] };
    final int length = leftData.length();
    if (length != rightData.length())
        throw new IllegalArgumentException();
    if (3 != dimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(dimensions));
    }
    for (int i = 1; i < inObj.length; i++) {
        if (Tensor.length(dimensions) != Tensor.length(inObj[i].getData().getDimensions())) {
            throw new IllegalArgumentException(Arrays.toString(dimensions) + " != " + Arrays.toString(inObj[i].getData().getDimensions()));
        }
    }
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    return new Result(CudaSystem.run(gpu -> {
        @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_ADD, precision);
        @Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
        // .getDenseAndFree(gpu);//.moveTo(gpu.getDeviceNumber());
        @Nullable final CudaTensor lPtr = gpu.getTensor(leftData, precision, MemoryType.Device, false);
        // .getDenseAndFree(gpu);//.moveTo(gpu.getDeviceNumber());
        @Nullable final CudaTensor rPtr = gpu.getTensor(rightData, precision, MemoryType.Device, false);
        @Nonnull final CudaMemory outputPtr = gpu.allocate(precision.size * Tensor.length(dimensions) * length, MemoryType.Managed, true);
        CudaMemory lPtrMemory = lPtr.getMemory(gpu);
        CudaMemory rPtrMemory = rPtr.getMemory(gpu);
        gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(leftFactor), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(rightFactor), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr());
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        lPtrMemory.dirty();
        rPtrMemory.dirty();
        outputPtr.dirty();
        rPtrMemory.freeRef();
        lPtrMemory.freeRef();
        CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, outputDescriptor, precision);
        Stream.<ReferenceCounting>of(opDescriptor, lPtr, rPtr).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
    }, leftData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        Runnable a = () -> {
            if (inObj[0].isAlive()) {
                CudaTensorList tensorList = CudaSystem.run(gpu -> {
                    @Nullable final CudaTensor lPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    @Nonnull final CudaMemory passbackPtr = gpu.allocate(precision.size * Tensor.length(dimensions) * length, MemoryType.Managed.normalize(), true);
                    @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
                    CudaMemory lPtrMemory = lPtr.getMemory(gpu);
                    gpu.cudnnTransformTensor(precision.getPointer(leftFactor), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(0.0), passbackDescriptor.getPtr(), passbackPtr.getPtr());
                    assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
                    passbackPtr.dirty();
                    lPtrMemory.dirty();
                    lPtrMemory.freeRef();
                    CudaTensor cudaTensor = CudaTensor.wrap(passbackPtr, passbackDescriptor, precision);
                    lPtr.freeRef();
                    return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
                }, delta);
                inObj[0].accumulate(buffer, tensorList);
            }
        };
        Runnable b = () -> {
            if (inObj[1].isAlive()) {
                CudaTensorList tensorList = CudaSystem.run(gpu -> {
                    @Nullable final CudaTensor lPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                    @Nonnull final CudaMemory outputPtr = gpu.allocate(precision.size * Tensor.length(dimensions) * length, MemoryType.Managed.normalize(), true);
                    @Nonnull final CudaDevice.CudaTensorDescriptor passbackDescriptor = gpu.newTensorDescriptor(precision, length, dimensions[2], dimensions[1], dimensions[0], dimensions[2] * dimensions[1] * dimensions[0], dimensions[1] * dimensions[0], dimensions[0], 1);
                    CudaMemory lPtrMemory = lPtr.getMemory(gpu);
                    gpu.cudnnTransformTensor(precision.getPointer(rightFactor), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(0.0), passbackDescriptor.getPtr(), outputPtr.getPtr());
                    outputPtr.dirty();
                    lPtrMemory.dirty();
                    lPtrMemory.freeRef();
                    CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, passbackDescriptor, precision);
                    lPtr.freeRef();
                    return CudaTensorList.wrap(cudaTensor, length, dimensions, precision);
                }, delta);
                inObj[1].accumulate(buffer, tensorList);
            }
        };
        if (CoreSettings.INSTANCE.isSingleThreaded())
            TestUtil.runAllSerial(a, b);
        else
            TestUtil.runAllParallel(a, b);
    }) {

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(x -> x.freeRef());
            leftData.freeRef();
            rightData.freeRef();
        }

        @Override
        public boolean isAlive() {
            for (@Nonnull final Result element : inObj) if (element.isAlive()) {
                return true;
            }
            return false;
        }
    };
}

Example 8 with DeltaSet

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

the class MaxImageBandLayer method eval.

@Nonnull
@Override
public Result eval(@Nonnull final Result... inObj) {
    assert 1 == inObj.length;
    final TensorList inputData = inObj[0].getData();
    inputData.addRef();
    inputData.length();
    @Nonnull final int[] inputDims = inputData.getDimensions();
    assert 3 == inputDims.length;
    Arrays.stream(inObj).forEach(nnResult -> nnResult.addRef());
    final Coordinate[][] maxCoords = inputData.stream().map(data -> {
        Coordinate[] coordinates = IntStream.range(0, inputDims[2]).mapToObj(band -> {
            return data.coordStream(true).filter(e -> e.getCoords()[2] == band).max(Comparator.comparing(c -> data.get(c))).get();
        }).toArray(i -> new Coordinate[i]);
        data.freeRef();
        return coordinates;
    }).toArray(i -> new Coordinate[i][]);
    return new Result(TensorArray.wrap(IntStream.range(0, inputData.length()).mapToObj(dataIndex -> {
        Tensor tensor = inputData.get(dataIndex);
        final DoubleStream doubleStream = IntStream.range(0, inputDims[2]).mapToDouble(band -> {
            final int[] maxCoord = maxCoords[dataIndex][band].getCoords();
            double v = tensor.get(maxCoord[0], maxCoord[1], band);
            return v;
        });
        Tensor tensor1 = new Tensor(1, 1, inputDims[2]).set(Tensor.getDoubles(doubleStream, inputDims[2]));
        tensor.freeRef();
        return tensor1;
    }).toArray(i -> new Tensor[i])), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (inObj[0].isAlive()) {
            @Nonnull TensorArray tensorArray = TensorArray.wrap(IntStream.range(0, delta.length()).parallel().mapToObj(dataIndex -> {
                Tensor deltaTensor = delta.get(dataIndex);
                @Nonnull final Tensor passback = new Tensor(inputData.getDimensions());
                IntStream.range(0, inputDims[2]).forEach(b -> {
                    final int[] maxCoord = maxCoords[dataIndex][b].getCoords();
                    passback.set(new int[] { maxCoord[0], maxCoord[1], b }, deltaTensor.get(0, 0, b));
                });
                deltaTensor.freeRef();
                return passback;
            }).toArray(i -> new Tensor[i]));
            inObj[0].accumulate(buffer, tensorArray);
        }
    }) {

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
            inputData.freeRef();
        }

        @Override
        public boolean isAlive() {
            return inObj[0].isAlive();
        }
    };
}

Example 9 with DeltaSet

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

the class MaxPoolingLayer method eval.

@Nonnull
@Override
public Result eval(@Nonnull final Result... inObj) {
    Arrays.stream(inObj).forEach(nnResult -> nnResult.addRef());
    final Result in = inObj[0];
    in.getData().length();
    @Nonnull final int[] inputDims = in.getData().getDimensions();
    final List<Tuple2<Integer, int[]>> regions = MaxPoolingLayer.calcRegionsCache.apply(new MaxPoolingLayer.CalcRegionsParameter(inputDims, kernelDims));
    final Tensor[] outputA = IntStream.range(0, in.getData().length()).mapToObj(dataIndex -> {
        final int[] newDims = IntStream.range(0, inputDims.length).map(i -> {
            return (int) Math.ceil(inputDims[i] * 1.0 / kernelDims[i]);
        }).toArray();
        @Nonnull final Tensor output = new Tensor(newDims);
        return output;
    }).toArray(i -> new Tensor[i]);
    Arrays.stream(outputA).mapToInt(x -> x.length()).sum();
    @Nonnull final int[][] gradientMapA = new int[in.getData().length()][];
    IntStream.range(0, in.getData().length()).forEach(dataIndex -> {
        @Nullable final Tensor input = in.getData().get(dataIndex);
        final Tensor output = outputA[dataIndex];
        @Nonnull final IntToDoubleFunction keyExtractor = inputCoords -> input.get(inputCoords);
        @Nonnull final int[] gradientMap = new int[input.length()];
        regions.parallelStream().forEach(tuple -> {
            final Integer from = tuple.getFirst();
            final int[] toList = tuple.getSecond();
            int toMax = -1;
            double bestValue = Double.NEGATIVE_INFINITY;
            for (final int c : toList) {
                final double value = keyExtractor.applyAsDouble(c);
                if (-1 == toMax || bestValue < value) {
                    bestValue = value;
                    toMax = c;
                }
            }
            gradientMap[from] = toMax;
            output.set(from, input.get(toMax));
        });
        input.freeRef();
        gradientMapA[dataIndex] = gradientMap;
    });
    return new Result(TensorArray.wrap(outputA), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList data) -> {
        if (in.isAlive()) {
            @Nonnull TensorArray tensorArray = TensorArray.wrap(IntStream.range(0, in.getData().length()).parallel().mapToObj(dataIndex -> {
                @Nonnull final Tensor backSignal = new Tensor(inputDims);
                final int[] ints = gradientMapA[dataIndex];
                @Nullable final Tensor datum = data.get(dataIndex);
                for (int i = 0; i < datum.length(); i++) {
                    backSignal.add(ints[i], datum.get(i));
                }
                datum.freeRef();
                return backSignal;
            }).toArray(i -> new Tensor[i]));
            in.accumulate(buffer, tensorArray);
        }
    }) {

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
        }

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

Example 10 with DeltaSet

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

the class MonitoringSynapse method eval.

@Override
public Result eval(@Nonnull final Result... inObj) {
    assert 1 == inObj.length;
    final Result input = inObj[0];
    final TensorList inputdata = input.getData();
    input.addRef();
    inputdata.addRef();
    System.nanoTime();
    System.nanoTime();
    totalBatches++;
    totalItems += inputdata.length();
    forwardStatistics.clear();
    inputdata.stream().parallel().forEach(t -> {
        forwardStatistics.add(t.getData());
        t.freeRef();
    });
    return new Result(inputdata, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList data) -> {
        backpropStatistics.clear();
        data.addRef();
        input.accumulate(buffer, data);
        data.stream().parallel().forEach(t -> {
            backpropStatistics.add(t.getData());
            t.freeRef();
        });
    }) {

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

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