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

Example 51 with DeltaSet

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

the class ImgCropLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    assert 1 == inObj.length;
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    assert 3 == inputData.getDimensions().length;
    final int length = inputData.length();
    @Nonnull int[] dimIn = inputData.getDimensions();
    if (dimIn[0] == sizeX && dimIn[1] == sizeY) {
        return input;
    }
    @Nonnull final int[] dimOut = Arrays.copyOf(dimIn, 3);
    dimOut[0] = sizeX;
    dimOut[1] = sizeY;
    final TensorList outputData = CudaSystem.run(gpu -> {
        @Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
        inputData.freeRef();
        boolean dirty = dimOut[0] <= dimIn[0] && dimOut[1] <= dimIn[1];
        assert dimOut[0] > 0;
        assert dimOut[1] > 0;
        assert dimOut[2] > 0;
        CudaTensor cudaTensor = copy(gpu, inputTensor, length, dimIn, dimOut, dirty, precision);
        Stream.<ReferenceCounting>of(inputTensor).forEach(ReferenceCounting::freeRef);
        return CudaTensorList.wrap(cudaTensor, length, dimOut, precision);
    }, inputData);
    return new Result(outputData, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        if (!Arrays.equals(delta.getDimensions(), outputData.getDimensions())) {
            throw new AssertionError(Arrays.toString(delta.getDimensions()) + " != " + Arrays.toString(outputData.getDimensions()));
        }
        if (delta.length() != outputData.length()) {
            throw new AssertionError(delta.length() + " != " + outputData.length());
        }
        assert delta.length() == length;
        if (input.isAlive()) {
            final TensorList passbackTensorList = CudaSystem.run(gpu -> {
                @Nullable final CudaTensor errorPtr = gpu.getTensor(delta, precision, MemoryType.Device, false);
                delta.freeRef();
                boolean dirty = dimOut[0] >= dimIn[0] && dimOut[1] >= dimIn[1];
                CudaTensor cudaTensor = copy(gpu, errorPtr, length, dimOut, dimIn, dirty, precision);
                Stream.<ReferenceCounting>of(errorPtr).forEach(ReferenceCounting::freeRef);
                return CudaTensorList.wrap(cudaTensor, length, dimIn, precision);
            }, delta);
            input.accumulate(buffer, passbackTensorList);
        } else {
            delta.freeRef();
        }
    }) {

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

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

        @Override
        public boolean isAlive() {
            return Arrays.stream(inObj).anyMatch(x -> x.isAlive());
        }
    };
}

Example 52 with DeltaSet

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

the class ImgTileAssemblyLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    if (1 == inObj.length) {
        return inObj[0];
    }
    int[] inputDimensions = inObj[0].getData().getDimensions();
    assert 3 == inputDimensions.length;
    final int length = inObj[0].getData().length();
    int[] outputDims = getOutputDims(inObj);
    final TensorList outputData = CudaSystem.run(gpu -> {
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        assert outputDims[0] > 0;
        assert outputDims[1] > 0;
        assert outputDims[2] > 0;
        @Nonnull final CudaMemory outputBuffer = gpu.allocate((long) length * outputDims[2] * outputDims[1] * outputDims[0] * precision.size, MemoryType.Managed.normalize(), false);
        int totalWidth = 0;
        int totalHeight = 0;
        int inputIndex = 0;
        List<CopyParams> copies = new ArrayList<>();
        for (int row = 0; row < rows; row++) {
            int positionX = 0;
            int rowHeight = 0;
            for (int col = 0; col < columns; col++) {
                int[] tileDimensions = inObj[inputIndex].getData().getDimensions();
                rowHeight = Math.max(rowHeight, tileDimensions[1]);
                copies.add(new CopyParams(gpu, inObj, outputBuffer, length, outputDims, tileDimensions, inputIndex, positionX, totalHeight));
                positionX += tileDimensions[0];
                inputIndex += 1;
                assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
            }
            totalHeight += rowHeight;
            totalWidth = Math.max(totalWidth, positionX);
        }
        assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
        Stream<CopyParams> stream = copies.stream();
        if (!CoreSettings.INSTANCE.isSingleThreaded() && parallel)
            stream = stream.parallel();
        stream.forEach(this::copy);
        Arrays.stream(inObj).forEach(r -> r.getData().freeRef());
        CudaDevice.CudaTensorDescriptor descriptor = gpu.newTensorDescriptor(precision, length, outputDims[2], outputDims[1], outputDims[0]);
        CudaTensor ptr = CudaTensor.wrap(outputBuffer, descriptor, precision);
        return CudaTensorList.wrap(ptr, length, outputDims, precision);
    }, Arrays.stream(inObj).map(Result::getData).toArray());
    return new Result(outputData, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList error) -> {
        if (!Arrays.equals(error.getDimensions(), outputData.getDimensions())) {
            throw new AssertionError(Arrays.toString(error.getDimensions()) + " != " + Arrays.toString(outputData.getDimensions()));
        }
        if (error.length() != outputData.length()) {
            throw new AssertionError(error.length() + " != " + outputData.length());
        }
        assert error.length() == length;
        int totalHeight = 0;
        int inputIndex = 0;
        List<BackpropParams> tasks = new ArrayList<>();
        for (int row = 0; row < rows; row++) {
            int positionX = 0;
            int rowHeight = 0;
            for (int col = 0; col < columns; col++) {
                Result in = inObj[inputIndex];
                int[] tileDimensions = in.getData().getDimensions();
                rowHeight = Math.max(rowHeight, tileDimensions[1]);
                if (inObj[inputIndex].isAlive()) {
                    tasks.add(new BackpropParams(inObj, buffer, error, outputDims, tileDimensions, length, positionX, totalHeight, inputIndex));
                }
                positionX += tileDimensions[0];
                inputIndex += 1;
            }
            totalHeight += rowHeight;
        }
        Stream<BackpropParams> stream = tasks.stream();
        if (!CoreSettings.INSTANCE.isSingleThreaded() && parallel)
            stream = stream.parallel();
        stream.forEach(this::backprop);
    }) {

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

        @Override
        public boolean isAlive() {
            return Arrays.stream(inObj).anyMatch(x -> x.isAlive());
        }
    };
}

Example 53 with DeltaSet

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

the class ImgTileSelectLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    assert 1 == inObj.length;
    final Result input = inObj[0];
    final TensorList inputData = input.getData();
    assert 3 == inputData.getDimensions().length;
    final int length = inputData.length();
    @Nonnull int[] dimIn = inputData.getDimensions();
    if (dimIn[0] == sizeY && dimIn[1] == sizeX) {
        return input;
    }
    @Nonnull final int[] dimOut = getViewDimensions(dimIn, new int[] { sizeY, sizeX, dimIn[2] }, new int[] { positionX, positionY, 0 });
    final TensorList outputData = CudaSystem.run(gpu -> {
        assert dimOut[0] > 0;
        assert dimOut[1] > 0;
        assert dimOut[2] > 0;
        boolean dirty = dimOut[0] == dimIn[0] && dimOut[1] == dimIn[1];
        CudaTensor cudaTensor = copy(gpu, inputData, dimIn, dimOut, precision, this.positionX, this.positionY, dirty);
        return CudaTensorList.wrap(cudaTensor, length, dimOut, precision);
    }, inputData);
    int[] outputDimensions = outputData.getDimensions();
    assert length == outputData.length();
    return new Result(outputData, (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList error) -> {
        if (!Arrays.equals(error.getDimensions(), outputDimensions)) {
            throw new AssertionError(Arrays.toString(error.getDimensions()) + " != " + Arrays.toString(outputDimensions));
        }
        if (error.length() != length) {
            throw new AssertionError(error.length() + " != " + length);
        }
        assert error.length() == inputData.length();
        if (input.isAlive()) {
            final TensorList passbackTensorList = CudaSystem.run(gpu -> {
                boolean dirty = dimOut[0] >= dimIn[0] && dimOut[1] >= dimIn[1];
                CudaTensor cudaTensor = copy(gpu, error, dimOut, dimIn, precision, -this.positionX, -this.positionY, dirty);
                return CudaTensorList.wrap(cudaTensor, length, dimIn, precision);
            }, error);
            input.accumulate(buffer, passbackTensorList);
        }
    }) {

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

        @Override
        public boolean isAlive() {
            return Arrays.stream(inObj).anyMatch(x -> x.isAlive());
        }
    };
}

Example 54 with DeltaSet

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

the class ImgTileSubnetLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    assert 1 == inObj.length;
    Result input = inObj[0];
    TensorList inputData = input.getData();
    @Nonnull final int[] inputDims = inputData.getDimensions();
    assert 3 == inputDims.length;
    int bands = inputDims[2];
    int length = inputData.length();
    CudaTensor passback = CudaSystem.run(gpu -> {
        return CudaTensor.wrap(gpu.allocate(inputData.getElements() * precision.size, MemoryType.Managed, true), gpu.newTensorDescriptor(precision, length, inputDims[2], inputDims[1], inputDims[0]), precision);
    });
    try {
        AtomicInteger counter = new AtomicInteger(0);
        int cols = (int) (Math.ceil((inputDims[0] - width) * 1.0 / strideX) + 1);
        int rows = (int) (Math.ceil((inputDims[1] - height) * 1.0 / strideY) + 1);
        if (cols == 1 && rows == 1)
            return getInner().evalAndFree(inObj);
        ArrayList<CudaTensor> tiles = new ArrayList<>();
        int[] tileDimensions = { width, height, bands };
        Result[][] tileResults = new Result[rows][];
        for (int row = 0; row < rows; row++) {
            tileResults[row] = new Result[cols];
            for (int col = 0; col < cols; col++) {
                int positionX = col * strideX;
                int positionY = row * strideY;
                assert positionX >= 0;
                assert positionY >= 0;
                assert positionX < inputDims[0];
                assert positionY < inputDims[1];
                CudaTensor tile = CudaSystem.run(gpu -> {
                    return ImgTileSelectLayer.copy(gpu, inputData, inputData.getDimensions(), tileDimensions, precision, positionX, positionY, true);
                });
                passback.addRef();
                tileResults[row][col] = getInner().evalAndFree(new Result(CudaTensorList.wrap(tile, length, tileDimensions, precision), (DeltaSet<Layer> ctx, TensorList delta) -> {
                    CudaSystem.run(gpu -> {
                        ImgTileSelectLayer.copy(gpu, delta, tileDimensions, -positionX, -positionY, precision, passback).freeRef();
                    });
                    if (counter.incrementAndGet() >= rows * cols) {
                        counter.set(0);
                        input.accumulate(ctx, CudaTensorList.create(passback, length, inputDims, precision));
                    }
                }) {

                    @Override
                    protected void _free() {
                        super._free();
                        passback.freeRef();
                    }
                });
            }
        }
        inputData.freeRef();
        logger.debug(String.format("Broke input %s into %s rows, %s cols", Arrays.toString(inputDims), rows, cols));
        Result result = new ImgTileAssemblyLayer(cols, rows).setParallel(parallel).setPrecision(precision).evalAndFree(Arrays.stream(tileResults).flatMap(Arrays::stream).toArray(i -> new Result[i]));
        return new Result(result.getData(), (ctx, delta) -> {
            result.accumulate(ctx, delta);
        }) {

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

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

Example 55 with DeltaSet

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

the class NProductLayer method evalAndFree.

@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
    if (!CudaSystem.isEnabled())
        return getCompatibilityLayer().evalAndFree(inObj);
    if (inObj.length <= 1) {
        throw new IllegalArgumentException("inObj.length=" + inObj.length);
    }
    @Nonnull final int[] dimensions = inObj[0].getData().getDimensions();
    final int length = inObj[0].getData().length();
    if (3 != dimensions.length) {
        throw new IllegalArgumentException("dimensions=" + Arrays.toString(dimensions));
    }
    for (int i = 1; i < inObj.length; i++) {
        TensorList data = inObj[i].getData();
        if (Tensor.length(dimensions) != Tensor.length(data.getDimensions())) {
            throw new IllegalArgumentException(Arrays.toString(dimensions) + " != " + Arrays.toString(data.getDimensions()));
        }
    }
    return new Result(CudaSystem.run(gpu -> {
        @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, 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);
        @Nonnull final TensorList result1 = Arrays.stream(inObj).map(x -> {
            TensorList data = x.getData();
            data.addRef();
            return data;
        }).reduce((l, r) -> {
            @Nullable final CudaTensor lPtr = gpu.getTensor(l, precision, MemoryType.Device, false);
            @Nullable final CudaTensor rPtr = gpu.getTensor(r, precision, MemoryType.Device, false);
            // assert lPtr.memory.size == rPtr.memory.size;
            @Nonnull final CudaMemory outputPtr = gpu.allocate((long) outputDescriptor.nStride * length * precision.size, MemoryType.Device, true);
            CudaMemory lPtrMemory = lPtr.getMemory(gpu);
            CudaMemory rPtrMemory = rPtr.getMemory(gpu);
            CudaSystem.handle(JCudnn.cudnnOpTensor(gpu.handle, opDescriptor.getPtr(), precision.getPointer(1.0), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(1.0), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
            lPtrMemory.dirty();
            rPtrMemory.dirty();
            outputPtr.dirty();
            lPtrMemory.freeRef();
            rPtrMemory.freeRef();
            Arrays.stream(new ReferenceCounting[] { lPtr, rPtr, l, r }).forEach(ReferenceCounting::freeRef);
            outputDescriptor.addRef();
            return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, dimensions, precision);
        }).get();
        Arrays.stream(new ReferenceCounting[] { opDescriptor, outputDescriptor }).forEach(ReferenceCounting::freeRef);
        return result1;
    }, Arrays.stream(inObj).map(Result::getData).toArray()), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
        for (int index = 0; index < inObj.length; index++) {
            final Result input = inObj[index];
            if (input.isAlive()) {
                final int _index = index;
                @Nonnull TensorList data = IntStream.range(0, inObj.length).mapToObj(i -> {
                    TensorList tensorList = i == _index ? delta : inObj[i].getData();
                    tensorList.addRef();
                    return tensorList;
                }).reduce((l, r) -> {
                    return CudaSystem.run(gpu -> {
                        @Nonnull final CudaResource<cudnnOpTensorDescriptor> opDescriptor = gpu.newOpDescriptor(cudnnOpTensorOp.CUDNN_OP_TENSOR_MUL, 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);
                        @Nullable final CudaTensor lPtr = gpu.getTensor(l, precision, MemoryType.Device, false);
                        @Nullable final CudaTensor rPtr = gpu.getTensor(r, precision, MemoryType.Device, false);
                        // assert lPtr.memory.size == rPtr.memory.size;
                        @Nonnull final CudaMemory outputPtr = gpu.allocate((long) outputDescriptor.nStride * length * precision.size, MemoryType.Device, true);
                        CudaMemory lPtrMemory = lPtr.getMemory(gpu);
                        CudaMemory rPtrMemory = rPtr.getMemory(gpu);
                        CudaSystem.handle(JCudnn.cudnnOpTensor(gpu.handle, opDescriptor.getPtr(), precision.getPointer(1.0), lPtr.descriptor.getPtr(), lPtrMemory.getPtr(), precision.getPointer(1.0), rPtr.descriptor.getPtr(), rPtrMemory.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
                        lPtrMemory.dirty();
                        rPtrMemory.dirty();
                        outputPtr.dirty();
                        lPtrMemory.freeRef();
                        rPtrMemory.freeRef();
                        Stream.of(lPtr, rPtr, opDescriptor, l, r).forEach(ReferenceCounting::freeRef);
                        return CudaTensorList.wrap(CudaTensor.wrap(outputPtr, outputDescriptor, precision), length, dimensions, precision);
                    }, l, r);
                }).get();
                input.accumulate(buffer, data);
            }
        }
        delta.freeRef();
    }) {

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

        @Override
        protected void _free() {
            Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
            for (int i = 0; i < inObj.length; i++) {
                inObj[i].getData().freeRef();
            }
        }

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