public static void SpatialMaxPoolingBackward(Tensor input, Tensor gradOutput, Tensor gradInput, Tensor indices, ConvolutionDesc2d cd, bool ceilMode)
{
int dimw = 3;
int dimh = 2;
int dimc = 1;
long nbatch = input.Sizes[0];
long nslices = input.Sizes[dimc];
long iheight = input.Sizes[dimh];
long iwidth = input.Sizes[dimw];
long owidth = gradOutput.Sizes[dimw];
long oheight = gradOutput.Sizes[dimh];
Ops.Fill(gradInput, 0);
using Tensor gradOutputContig = Ops.AsContiguous(gradOutput);
for (int i = 0; i < nbatch; ++i)
{
using Tensor gradInput_i = gradInput.Select(0, i);
using Tensor gradOutput_i = gradOutputContig.Select(0, i);
using Tensor indices_i = indices.Select(0, i);
using (NativeWrapper.BuildTensorRefPtr(gradInput_i, out IntPtr gradInput_iPtr))
using (NativeWrapper.BuildTensorRefPtr(gradOutput_i, out IntPtr gradOutput_iPtr))
using (NativeWrapper.BuildTensorRefPtr(indices_i, out IntPtr indices_iPtr))
{
CpuOpsNative.TS_SpatialMaxPooling_updateGradInput_frame(gradInput_iPtr, gradOutput_iPtr, indices_iPtr,
nslices, iwidth, iheight,
owidth, oheight,
cd.dW, cd.dH);
}
}
}
private static void Conv2BackwardInputFrame(Tensor gradOutput, Tensor gradInput, Tensor weight, Tensor fgradInput, ConvolutionDesc2d cd)
{
using (Tensor gradOutput2d = gradOutput.View(gradOutput.Sizes[0], gradOutput.Sizes[1] * gradOutput.Sizes[2]))
{
Ops.Addmm(fgradInput, 0, fgradInput, 1, weight, gradOutput2d);
}
Ops.Fill(gradInput, 0);
using (NativeWrapper.BuildTensorRefPtr(fgradInput, out IntPtr fgradInputPtr))
using (NativeWrapper.BuildTensorRefPtr(gradInput, out IntPtr gradInputPtr))
{
CpuOpsNative.TS_Unfolded_Acc(fgradInputPtr, gradInputPtr, cd.kW, cd.kH, cd.dW, cd.dH, cd.padW, cd.padH,
(int)gradInput.Sizes[0], (int)gradInput.Sizes[2], (int)gradInput.Sizes[1],
(int)gradOutput.Sizes[2], (int)gradOutput.Sizes[1]);
}
}
/// <summary>
/// Spatials the maximum pooling backward.
/// </summary>
/// <param name="input">The input.</param>
/// <param name="gradOutput">The grad output.</param>
/// <param name="gradInput">The grad input.</param>
/// <param name="indices">The indices.</param>
/// <param name="cd">The cd.</param>
/// <param name="ceilMode">if set to <c>true</c> [ceil mode].</param>
public static void SpatialMaxPoolingBackward(NDArray input, NDArray gradOutput, NDArray gradInput, NDArray indices, ConvolutionDesc2d cd, bool ceilMode)
{
var dimw = 3;
var dimh = 2;
var dimc = 1;
var nbatch = input.Shape[0];
var nslices = input.Shape[dimc];
var iheight = input.Shape[dimh];
var iwidth = input.Shape[dimw];
var owidth = gradOutput.Shape[dimw];
var oheight = gradOutput.Shape[dimh];
Ops.Fill(gradInput, 0);
using (var gradOutputContig = Ops.AsContiguous(gradOutput))
{
for (int i = 0; i < nbatch; ++i)
{
using (var gradInput_i = gradInput.Select(0, i))
using (var gradOutput_i = gradOutputContig.Select(0, i))
using (var indices_i = indices.Select(0, i))
{
IntPtr gradInput_iPtr, gradOutput_iPtr, indices_iPtr;
using (NativeWrapper.BuildTensorRefPtr(gradInput_i, out gradInput_iPtr))
using (NativeWrapper.BuildTensorRefPtr(gradOutput_i, out gradOutput_iPtr))
using (NativeWrapper.BuildTensorRefPtr(indices_i, out indices_iPtr))
{
CpuOpsNative.TS_SpatialMaxPooling_updateGradInput_frame(gradInput_iPtr, gradOutput_iPtr, indices_iPtr,
nslices, iwidth, iheight,
owidth, oheight,
cd.dW, cd.dH);
}
}
}
}
}
public static void SpatialMaxPoolingForward(Tensor input, Tensor output, Tensor indices, ConvolutionDesc2d cd, bool ceilMode)
{
if (input.DimensionCount != 4)
{
throw new ArgumentException("input must be a 4D tensor");
}
var dimw = 3;
var dimh = 2;
var dimc = 1;
if (input.Sizes[dimw] < cd.kW - cd.padW || input.Sizes[dimh] < cd.kH - cd.padH)
{
throw new InvalidOperationException("input image is smaller than kernel size");
}
if (cd.padW > cd.kW / 2 || cd.padH > cd.kH / 2)
{
throw new InvalidOperationException("pad should be smaller than half of the kernel size");
}
var nbatch = input.Sizes[0];
var nslices = input.Sizes[dimc];
var iheight = input.Sizes[dimh];
var iwidth = input.Sizes[dimw];
long owidth;
long oheight;
if (ceilMode)
{
// ReSharper disable once ArrangeRedundantParentheses
oheight = (long)(Math.Ceiling((float)(iheight - cd.kH + 2 * cd.padH) / cd.dH)) + 1;
// ReSharper disable once ArrangeRedundantParentheses
owidth = (long)(Math.Ceiling((float)(iwidth - cd.kW + 2 * cd.padW) / cd.dW)) + 1;
}
else
{
// ReSharper disable once ArrangeRedundantParentheses
oheight = (long)(Math.Floor((float)(iheight - cd.kH + 2 * cd.padH) / cd.dH)) + 1;
// ReSharper disable once ArrangeRedundantParentheses
owidth = (long)(Math.Floor((float)(iwidth - cd.kW + 2 * cd.padW) / cd.dW)) + 1;
}
if (cd.padW != 0 || cd.padH != 0)
{
// ensure that the last pooling starts inside the image
if ((oheight - 1) * cd.dH >= iheight + cd.padH)
{
--oheight;
}
if ((owidth - 1) * cd.dW >= iwidth + cd.padW)
{
--owidth;
}
}
using var inputContig = Ops.AsContiguous(input);
for (var i = 0; i < nbatch; ++i)
{
using var input_i = inputContig.Select(0, i);
using var output_i = output.Select(0, i);
using var indices_i = indices.Select(0, i);
using (NativeWrapper.BuildTensorRefPtr(input_i, out var input_iPtr))
{
using (NativeWrapper.BuildTensorRefPtr(output_i, out var output_iPtr))
{
using (NativeWrapper.BuildTensorRefPtr(indices_i, out var indices_iPtr))
{
CpuOpsNative.TS_SpatialMaxPooling_updateOutput_frame(input_iPtr, output_iPtr, indices_iPtr,
nslices, iwidth, iheight,
owidth, oheight,
cd.kW, cd.kH, cd.dW, cd.dH, cd.padW, cd.padH);
}
}
}
}
}