use of com.simiacryptus.mindseye.lang.DeltaSet in project MindsEye by SimiaCryptus.
the class GramianLayer method evalAndFree.
@Nullable
@Override
public Result evalAndFree(final Result... inObj) {
assert 1 == inObj.length;
TensorList inputData = inObj[0].getData();
int[] inputDimensions = inputData.getDimensions();
assert 3 == inputDimensions.length;
return new Result(CudaSystem.run(gpu -> {
CudaTensor tensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
CudaTensorList output = getOutput(gpu, tensor);
tensor.freeRef();
return output;
}, inputData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
@Nonnull final int[] outputDimensions = { 1, 1, inputDimensions[2] * inputDimensions[2] };
if (!Arrays.equals(delta.getDimensions(), outputDimensions)) {
throw new AssertionError(Arrays.toString(delta.getDimensions()) + " != " + Arrays.toString(outputDimensions));
}
if (inObj[0].isAlive()) {
final TensorList passbackTensorList = CudaSystem.run(gpu -> {
@Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
delta.freeRef();
CudaTensorList feedback = getFeedback(gpu, inputTensor, deltaTensor);
deltaTensor.freeRef();
inputTensor.freeRef();
return feedback;
}, delta);
inObj[0].accumulate(buffer, passbackTensorList);
} else {
delta.freeRef();
}
}) {
@Override
public final void accumulate(DeltaSet<Layer> buffer, TensorList delta) {
getAccumulator().accept(buffer, delta);
}
@Override
protected void _free() {
inputData.freeRef();
Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
}
@Override
public boolean isAlive() {
return Arrays.stream(inObj).anyMatch(x -> x.isAlive());
}
};
}
use of com.simiacryptus.mindseye.lang.DeltaSet in project MindsEye by SimiaCryptus.
the class ImgBandBiasLayer 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);
}
Result input = inObj[0];
final TensorList leftData = input.getData();
@Nonnull final int[] inputDimensions = leftData.getDimensions();
final int length = leftData.length();
if (3 != inputDimensions.length) {
throw new IllegalArgumentException("dimensions=" + Arrays.toString(inputDimensions));
}
// assert !right.isAlive();
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, inputDimensions[2], inputDimensions[1], inputDimensions[0], inputDimensions[2] * inputDimensions[1] * inputDimensions[0], inputDimensions[1] * inputDimensions[0], inputDimensions[0], 1);
@Nullable final CudaTensor inputTensor = gpu.getTensor(leftData, precision, MemoryType.Device, false);
CudaMemory biasMem = gpu.allocate(bias.length() * precision.size, MemoryType.Device, true).write(precision, bias.getData());
int[] biasDim = bias.getDimensions();
CudaDevice.CudaTensorDescriptor biasDescriptor = gpu.newTensorDescriptor(precision, 1, biasDim[2], biasDim[1], biasDim[0], biasDim[2] * biasDim[1] * biasDim[0], biasDim[1] * biasDim[0], biasDim[0], 1);
// assert lPtr.size == rPtr.size;
@Nonnull final CudaMemory outputPtr = gpu.allocate((long) precision.size * outputDescriptor.nStride * length, MemoryType.Managed.normalize(), true);
CudaMemory inputMemory = inputTensor.getMemory(gpu);
CudaSystem.handle(gpu.cudnnOpTensor(opDescriptor.getPtr(), precision.getPointer(1.0), inputTensor.descriptor.getPtr(), inputMemory.getPtr(), precision.getPointer(1.0), biasDescriptor.getPtr(), biasMem.getPtr(), precision.getPointer(0.0), outputDescriptor.getPtr(), outputPtr.getPtr()));
assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
inputMemory.dirty();
biasMem.dirty();
outputPtr.dirty();
inputMemory.freeRef();
biasMem.freeRef();
biasDescriptor.freeRef();
inputTensor.freeRef();
opDescriptor.freeRef();
CudaTensor cudaTensor = CudaTensor.wrap(outputPtr, outputDescriptor, precision);
return CudaTensorList.wrap(cudaTensor, length, inputDimensions, precision);
}, leftData), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
if (!isFrozen()) {
@Nonnull double[] biasDelta = CudaSystem.run(gpu -> {
@Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, false);
CudaMemory biasMem = gpu.allocate(bias.length() * precision.size, MemoryType.Device, true).write(precision, bias.getData());
int[] biasDim = bias.getDimensions();
CudaDevice.CudaTensorDescriptor biasDescriptor = gpu.newTensorDescriptor(precision, 1, biasDim[2], biasDim[1], biasDim[0], biasDim[2] * biasDim[1] * biasDim[0], biasDim[1] * biasDim[0], biasDim[0], 1);
CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
gpu.cudnnConvolutionBackwardBias(precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), precision.getPointer(0.0), biasDescriptor.getPtr(), biasMem.getPtr());
assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
biasMem.dirty();
double[] biasV = new double[bias.length()];
biasMem.read(precision, biasV);
Stream.<ReferenceCounting>of(biasMem, deltaTensorMemory, deltaTensor, biasDescriptor).forEach(ReferenceCounting::freeRef);
return biasV;
}, delta);
buffer.get(ImgBandBiasLayer.this, bias).addInPlace(biasDelta).freeRef();
}
if (input.isAlive()) {
input.accumulate(buffer, delta);
} else {
delta.freeRef();
}
}) {
@Override
public final void accumulate(DeltaSet<Layer> buffer, TensorList delta) {
getAccumulator().accept(buffer, delta);
}
@Override
protected void _free() {
leftData.freeRef();
input.freeRef();
}
@Override
public boolean isAlive() {
for (@Nonnull final Result element : inObj) if (element.isAlive()) {
return true;
}
return false;
}
};
}
use of com.simiacryptus.mindseye.lang.DeltaSet in project MindsEye by SimiaCryptus.
the class ImgLinearSubnetLayer 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 length = inputData.length();
int maxBand = legs.stream().mapToInt(x -> x.toBand).max().getAsInt();
assert maxBand == inputDims[2] : maxBand + " != " + inputDims[2];
assert IntStream.range(0, maxBand).allMatch(i -> 1 == legs.stream().filter(x -> x.fromBand <= i && x.toBand > i).count());
CudaTensor passback = CudaSystem.run(gpu -> {
return CudaTensor.wrap(gpu.allocate(inputData.getElements() * precision.size, MemoryType.Device, true), gpu.newTensorDescriptor(precision, length, inputDims[2], inputDims[1], inputDims[0]), precision);
});
try {
AtomicInteger counter = new AtomicInteger(0);
SumInputsLayer sumInputsLayer = new SumInputsLayer();
try {
Result[] legResults = legs.stream().map(leg -> {
passback.addRef();
ImgBandSelectLayer imgBandSelectLayer = new ImgBandSelectLayer(leg.fromBand, leg.toBand);
input.addRef();
TensorList legData = imgBandSelectLayer.eval(input).getDataAndFree();
imgBandSelectLayer.freeRef();
return leg.inner.evalAndFree(new Result(legData, (DeltaSet<Layer> ctx, TensorList delta) -> {
int[] outputDimensions = delta.getDimensions();
int[] inputDimensions = inputDims;
synchronized (passback) {
CudaSystem.run(gpu -> {
@Nonnull final CudaDevice.CudaTensorDescriptor viewDescriptor = gpu.newTensorDescriptor(//
precision, //
length, //
outputDimensions[2], //
outputDimensions[1], //
outputDimensions[0], //
inputDimensions[2] * inputDimensions[1] * inputDimensions[0], //
inputDimensions[1] * inputDimensions[0], //
inputDimensions[0], 1);
final int byteOffset = viewDescriptor.cStride * leg.fromBand * precision.size;
assert delta.length() == inputData.length();
assert passback.getDeviceId() == gpu.getDeviceId();
// assert error.stream().flatMapToDouble(x-> Arrays.stream(x.getData())).allMatch(Double::isFinite);
@Nullable final CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
@Nonnull final CudaMemory passbackBuffer = passback.getMemory(gpu);
CudaMemory errorPtrMemory = deltaTensor.getMemory(gpu);
passbackBuffer.synchronize();
gpu.cudnnTransformTensor(precision.getPointer(1.0), deltaTensor.descriptor.getPtr(), errorPtrMemory.getPtr(), precision.getPointer(0.0), viewDescriptor.getPtr(), passbackBuffer.getPtr().withByteOffset(byteOffset));
errorPtrMemory.dirty();
passbackBuffer.dirty();
Stream.<ReferenceCounting>of(deltaTensor, viewDescriptor, passbackBuffer, errorPtrMemory).forEach(ReferenceCounting::freeRef);
}, passback);
}
if (counter.incrementAndGet() >= legs.size()) {
counter.set(0);
input.accumulate(ctx, CudaTensorList.create(passback, length, inputDims, precision));
}
}) {
@Override
protected void _free() {
super._free();
input.freeRef();
passback.freeRef();
}
});
}).toArray(i -> new Result[i]);
return sumInputsLayer.setParallel(parallel).setPrecision(precision).evalAndFree(legResults);
} finally {
sumInputsLayer.freeRef();
input.freeRef();
}
} finally {
passback.freeRef();
}
}
use of com.simiacryptus.mindseye.lang.DeltaSet in project MindsEye by SimiaCryptus.
the class ConvolutionLayer method eval.
@Nonnull
@Override
public Result eval(@Nonnull final Result... inObj) {
Arrays.stream(inObj).forEach(nnResult -> nnResult.addRef());
final Result input = inObj[0];
final TensorList batch = input.getData();
batch.addRef();
@Nonnull final int[] inputDims = batch.get(0).getDimensions();
@Nonnull final int[] kernelDims = kernel.getDimensions();
@Nullable final double[] kernelData = ConvolutionLayer.this.kernel.getData();
@Nonnull final ConvolutionController convolutionController = new ConvolutionController(inputDims, kernelDims, paddingX, paddingY);
final Tensor[] output = IntStream.range(0, batch.length()).mapToObj(dataIndex -> new Tensor(convolutionController.getOutputDims())).toArray(i -> new Tensor[i]);
try {
final double[][] inputBuffers = batch.stream().map(x -> {
@Nullable double[] data = x.getData();
x.detach();
return data;
}).toArray(i -> new double[i][]);
final double[][] outputBuffers = Arrays.stream(output).map(x -> x.getData()).toArray(i -> new double[i][]);
convolutionController.convolve(inputBuffers, kernelData, outputBuffers);
} catch (@Nonnull final Throwable e) {
throw new RuntimeException("Error mapCoords image res " + Arrays.toString(inputDims), e);
}
int outputLength = output.length;
return new Result(TensorArray.wrap(output), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList error) -> {
if (!isFrozen()) {
final double[][] inputBuffers = batch.stream().map(x -> {
@Nullable double[] data = x.getData();
x.freeRef();
return data;
}).toArray(i -> new double[i][]);
final double[][] outputBuffers = error.stream().map(x -> {
@Nullable double[] data = x.getData();
x.freeRef();
return data;
}).toArray(i -> new double[i][]);
@Nonnull final Tensor weightGradient = new Tensor(kernelDims);
convolutionController.gradient(inputBuffers, weightGradient.getData(), outputBuffers);
buffer.get(ConvolutionLayer.this, kernelData).addInPlace(weightGradient.getData()).freeRef();
weightGradient.freeRef();
}
if (input.isAlive()) {
final Tensor[] inputBufferTensors = IntStream.range(0, outputLength).mapToObj(dataIndex -> new Tensor(inputDims)).toArray(i -> new Tensor[i]);
final double[][] inputBuffers = Arrays.stream(inputBufferTensors).map(x -> {
@Nullable double[] data = x.getData();
return data;
}).toArray(i -> new double[i][]);
final double[][] outputBuffers = error.stream().map(x -> {
@Nullable double[] data = x.getData();
x.freeRef();
return data;
}).toArray(i -> new double[i][]);
convolutionController.backprop(inputBuffers, kernelData, outputBuffers);
@Nonnull TensorArray tensorArray = TensorArray.wrap(inputBufferTensors);
input.accumulate(buffer, tensorArray);
}
}) {
@Override
protected void _free() {
Arrays.stream(inObj).forEach(nnResult -> nnResult.freeRef());
batch.freeRef();
}
@Override
public boolean isAlive() {
return input.isAlive() || !isFrozen();
}
};
}
use of com.simiacryptus.mindseye.lang.DeltaSet in project MindsEye by SimiaCryptus.
the class ActivationLayer method evalAndFree.
@Nullable
@Override
public Result evalAndFree(@Nonnull final Result... inObj) {
if (!CudaSystem.isEnabled())
return getCompatibilityLayer().evalAndFree(inObj);
// assert Arrays.stream(inObj).flatMapToDouble(input->input.data.stream().flatMapToDouble(x-> Arrays.stream(x.getData()))).allMatch(v->Double.isFinite(v));
final Result inputResult = inObj[0];
final TensorList inputData = inputResult.getData();
@Nonnull final int[] inputSize = inputData.getDimensions();
@Nonnull final int[] outputSize = inputSize;
final int length = inputData.length();
final int inputDims = Tensor.length(inputSize);
try {
final CudaTensor outPtr = CudaSystem.run(gpu -> {
@Nullable final CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, false);
final CudaTensor outputTensor;
if (1 == inputData.currentRefCount() && 1 == inputTensor.currentRefCount() && (!inputResult.isAlive() || mode == Mode.RELU.id)) {
inputTensor.addRef();
outputTensor = inputTensor;
} else {
@Nonnull final CudaDevice.CudaTensorDescriptor outputDescriptor = gpu.newTensorDescriptor(precision, length, inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1);
@Nonnull final CudaMemory outputData = gpu.allocate((long) precision.size * inputDims * length, MemoryType.Managed.normalize(), true);
outputTensor = CudaTensor.wrap(outputData, outputDescriptor, precision);
}
@Nonnull final CudaResource<cudnnActivationDescriptor> activationDesc = gpu.newActivationDescriptor(mode, cudnnNanPropagation.CUDNN_NOT_PROPAGATE_NAN, 0);
try {
CudaMemory memory = inputTensor.getMemory(gpu);
CudaMemory tensorMemory = outputTensor.getMemory(gpu);
CudaSystem.handle(gpu.cudnnActivationForward(activationDesc.getPtr(), precision.getPointer(1.0), inputTensor.descriptor.getPtr(), memory.getPtr(), precision.getPointer(0.0), outputTensor.descriptor.getPtr(), tensorMemory.getPtr()));
assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
memory.dirty();
tensorMemory.dirty();
tensorMemory.freeRef();
memory.freeRef();
return outputTensor;
} catch (@Nonnull final Throwable e) {
throw new ComponentException("Error apply " + Arrays.toString(inputSize), e);
} finally {
activationDesc.freeRef();
inputTensor.freeRef();
}
}, inputData);
return new Result(CudaTensorList.create(outPtr, length, outputSize, precision), (@Nonnull final DeltaSet<Layer> buffer, @Nonnull final TensorList delta) -> {
if (inputResult.isAlive()) {
final TensorList data = CudaSystem.run(gpu -> {
@Nullable CudaTensor inputTensor = gpu.getTensor(inputData, precision, MemoryType.Device, true);
@Nullable CudaTensor deltaTensor = gpu.getTensor(delta, precision, MemoryType.Device, true);
assert length == delta.length();
CudaTensor localOut = outPtr.getDense(gpu);
delta.freeRef();
CudaTensor passbackTensor;
// if (sameStrides(deltaTensor.descriptor, inputTensor.descriptor)) {
// passbackTensor = deltaTensor;
// passbackTensor.addRef();
// }
// else {
// passbackTensor = deltaTensor.getDense(gpu);
// inputTensor = inputTensor.getDenseAndFree(gpu);
// }
passbackTensor = CudaTensor.wrap(gpu.allocate((long) Tensor.length(inputSize) * length * precision.size, MemoryType.Managed.normalize(), false), gpu.newTensorDescriptor(precision, length, inputSize[2], inputSize[1], inputSize[0], inputSize[2] * inputSize[1] * inputSize[0], inputSize[1] * inputSize[0], inputSize[0], 1), precision);
@Nonnull final CudaResource<cudnnActivationDescriptor> activationDesc = gpu.newActivationDescriptor(mode, cudnnNanPropagation.CUDNN_NOT_PROPAGATE_NAN, 0);
try {
CudaMemory localOutMemory = localOut.getMemory(gpu);
CudaMemory deltaTensorMemory = deltaTensor.getMemory(gpu);
CudaMemory inputTensorMemory = inputTensor.getMemory(gpu);
CudaMemory passbackTensorMemory = passbackTensor.getMemory(gpu);
CudaSystem.handle(gpu.cudnnActivationBackward(activationDesc.getPtr(), precision.getPointer(1.0), localOut.descriptor.getPtr(), localOutMemory.getPtr(), deltaTensor.descriptor.getPtr(), deltaTensorMemory.getPtr(), inputTensor.descriptor.getPtr(), inputTensorMemory.getPtr(), precision.getPointer(0.0), passbackTensor.descriptor.getPtr(), passbackTensorMemory.getPtr()));
assert CudaDevice.isThreadDeviceId(gpu.getDeviceId());
localOutMemory.dirty();
deltaTensorMemory.dirty();
inputTensorMemory.dirty();
passbackTensorMemory.dirty();
localOutMemory.freeRef();
deltaTensorMemory.freeRef();
inputTensorMemory.freeRef();
passbackTensorMemory.freeRef();
} catch (@Nonnull final Throwable e) {
throw new ComponentException("Error apply " + Arrays.toString(inputSize), e);
} finally {
localOut.freeRef();
inputTensor.freeRef();
deltaTensor.freeRef();
activationDesc.freeRef();
}
return CudaTensorList.wrap(passbackTensor, length, inputSize, precision);
}, delta);
inputResult.accumulate(buffer, data);
} else {
delta.freeRef();
}
}) {
@Override
public final void accumulate(DeltaSet<Layer> buffer, TensorList delta) {
getAccumulator().accept(buffer, delta);
}
@Override
protected void _free() {
inputData.freeRef();
outPtr.freeRef();
inputResult.freeRef();
}
@Override
public boolean isAlive() {
return inputResult.isAlive() || !isFrozen();
}
};
} catch (@Nonnull final Throwable e) {
throw new ComponentException("Error apply image res " + Arrays.toString(inputSize), e);
}
}
Aggregations