public static void PoolingBackward(DNNPoolingDesc desc, Tensor x, Tensor y, Tensor dx, Tensor dy)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var poolingDesc = new PoolingDescriptor();
poolingDesc.SetPoolingNdDescriptor((cudnnPoolingMode)desc.Mode, cudnnNanPropagation.PropagateNan, desc.WindowDims.Length,
desc.WindowDims, desc.Padding, desc.Strides);
using (var xPtr = GetDeviceVar(x))
using (var yPtr = GetDeviceVar(y))
using (var dxPtr = GetDeviceVar(dx))
using (var dyPtr = GetDeviceVar(dy))
using (var xDesc = GetDescriptor(x))
using (var yDesc = GetDescriptor(y))
using (var dxDesc = GetDescriptor(dx))
using (var dyDesc = GetDescriptor(dy))
{
// Note: ManagedCUDA argument names may be slightly misleading (src refers to 'y' here, and dest to 'x')
dnn.Value.PoolingBackward(poolingDesc, 1,
yDesc, yPtr,
dyDesc, dyPtr,
xDesc, xPtr,
0,
dxDesc, dxPtr);
}
}
}
public override void DoPoolGradient(Volume <double> input, Volume <double> outputGradient,
Volume <double> inputGradient, int windowWidth, int windowHeight,
int horizontalPad, int verticalPad, int horizontalStride, int verticalStride)
{
var inputStorage = input.Storage as VolumeStorage;
var inputGradientStorage = inputGradient.Storage as VolumeStorage;
var outputStorage = this._volumeStorage;
var outputGradientStorage = outputGradient.Storage as VolumeStorage;
// Copy to device if not already done
//outputStorage.CopyToDevice();
outputGradientStorage.CopyToDevice();
inputStorage.CopyToDevice();
inputGradientStorage.CopyToDevice();
// Synchro
this._context.DefaultStream.Synchronize();
using (var poolingDesc = new PoolingDescriptor())
using (var srcDesc = new TensorDescriptor())
using (var srcDiffDesc = new TensorDescriptor())
using (var destDesc = new TensorDescriptor())
using (var destDiffDesc = new TensorDescriptor())
{
var n = this.Shape.GetDimension(3);
var c = this.Shape.GetDimension(2);
var h = this.Shape.GetDimension(1);
var w = this.Shape.GetDimension(0);
srcDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Double, n, c, h, w);
srcDiffDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Double, n, c, h, w);
destDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Double,
inputStorage.Shape.GetDimension(3), inputStorage.Shape.GetDimension(2),
inputStorage.Shape.GetDimension(1), inputStorage.Shape.GetDimension(0));
destDiffDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Double,
inputStorage.Shape.GetDimension(3), inputStorage.Shape.GetDimension(2),
inputStorage.Shape.GetDimension(1), inputStorage.Shape.GetDimension(0));
poolingDesc.SetPooling2dDescriptor(cudnnPoolingMode.Max, cudnnNanPropagation.NotPropagateNan,
windowHeight, windowWidth,
verticalPad, horizontalPad, verticalStride, horizontalStride);
this._context.CudnnContext.PoolingBackward(poolingDesc, 1.0,
srcDesc, outputStorage.DeviceBuffer,
srcDiffDesc, outputGradientStorage.DeviceBuffer,
destDesc, inputStorage.DeviceBuffer,
0.0,
destDiffDesc, inputGradientStorage.DeviceBuffer);
}
inputGradientStorage.CopiedToDevice = true;
}
public override void Pool(Tensor t, int filterSize, int stride, Tensor.PoolType type, int paddingX, int paddingY, Tensor result)
{
t.CopyToDevice();
result.CopyToDevice();
using (var poolingDesc = new PoolingDescriptor())
using (var tDesc = new TensorDescriptor())
using (var resultDesc = new TensorDescriptor())
{
poolingDesc.SetPooling2dDescriptor(TensorPoolTypeToCuDNNPoolType(type), cudnnNanPropagation.NotPropagateNan, filterSize, filterSize, paddingX, paddingY, stride, stride);
tDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Float, t.Shape.Dimensions[3], t.Shape.Dimensions[2], t.Shape.Dimensions[1], t.Shape.Dimensions[0]);
resultDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Float, result.Shape.Dimensions[3], result.Shape.Dimensions[2], result.Shape.Dimensions[1], result.Shape.Dimensions[0]);
_CudnnContext.PoolingForward(poolingDesc, 1.0f, tDesc, t.GpuData.DeviceVar, 0.0f, resultDesc, result.GpuData.DeviceVar);
}
}
public static void PoolingForward(DNNPoolingDesc desc, Tensor x, Tensor y)
{
using (var dnn = CudaHelpers.TSContextForTensor(x).DNNForTensor(x))
{
var poolingDesc = new PoolingDescriptor();
poolingDesc.SetPoolingNdDescriptor((cudnnPoolingMode)desc.Mode, cudnnNanPropagation.PropagateNan, desc.WindowDims.Length,
desc.WindowDims, desc.Padding, desc.Strides);
using (var xPtr = GetDeviceVar(x))
using (var yPtr = GetDeviceVar(y))
using (var xDesc = GetDescriptor(x))
using (var yDesc = GetDescriptor(y))
{
dnn.Value.PoolingForward(poolingDesc, 1,
xDesc, xPtr,
0,
yDesc, yPtr);
}
}
}
public override void DoPool(Volume <double> result, int windowWidth, int windowHeight,
int horizontalPad, int verticalPad, int horizontalStride, int verticalStride)
{
var resultStorage = result.Storage as VolumeStorage;
if (resultStorage == null)
{
throw new ArgumentException($"{nameof(result)} storage should be VolumeStorage", nameof(result));
}
// Copy to device if not already done
this._volumeStorage.CopyToDevice();
resultStorage.CopyToDevice();
// Synchro
this._context.DefaultStream.Synchronize();
// Relu
using (var poolingDesc = new PoolingDescriptor())
using (var srcDesc = new TensorDescriptor())
using (var resultDesc = new TensorDescriptor())
{
srcDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Double,
this.Shape.GetDimension(3), this.Shape.GetDimension(2),
this.Shape.GetDimension(1), this.Shape.GetDimension(0));
resultDesc.SetTensor4dDescriptor(cudnnTensorFormat.NCHW, cudnnDataType.Double,
result.Shape.GetDimension(3), result.Shape.GetDimension(2),
result.Shape.GetDimension(1), result.Shape.GetDimension(0));
poolingDesc.SetPooling2dDescriptor(cudnnPoolingMode.Max, cudnnNanPropagation.NotPropagateNan,
windowHeight, windowWidth,
verticalPad, horizontalPad, verticalStride, horizontalStride);
this._context.CudnnContext.PoolingForward(poolingDesc, 1.0, srcDesc, this._volumeStorage.DeviceBuffer, 0.0,
resultDesc, resultStorage.DeviceBuffer);
}
resultStorage.CopiedToDevice = true;
}
public Pooling2D(Variable <T> data, PoolingMode mode, int kernelH, int kernelW, int strideH, int strideW)
{
Descriptor = new PoolingDescriptor();
Descriptor.Set2D(mode, NanPropagation.NOT_PROPAGATE_NAN, kernelH, kernelW, 0, 0, strideH, strideW);
var dataType = Dnn.DataTypeOf <T>();
var dataDesc = new TensorDescriptor();
dataDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, 10, (int)data.Shape[1], (int)data.Shape[2], (int)data.Shape[3]);
int n, c, h, w;
Descriptor.Get2dForwardOutputDim(dataDesc, out n, out c, out h, out w);
Data = data;
Output = Variable <T>(PartialShape.Create(-1, c, h, w));
AddInput(Data);
AddOutput(Output);
dataDesc.Dispose();
}
