Where is the C/C++ code for torch.nn.functional.fold()?

zfzhang · July 11, 2021, 3:05am

Hi,
I’m new to Pytorch C/C++ extention. My question is naive: Where is the C/C++ code for torch.nn.functional.fold()?

Thank you!

zfzhang · July 11, 2021, 5:46am

Ok, I’ve found it.

pytorch/pytorch/blob/c371542efc31b1abfe6f388042aa3ab0cef935f2/torch/csrc/api/include/torch/nn/functional/fold.h

#pragma once

#include <torch/nn/options/fold.h>

namespace torch {
namespace nn {
namespace functional {

#ifndef DOXYGEN_SHOULD_SKIP_THIS
namespace detail {
inline Tensor fold(const Tensor& input,
                   ExpandingArray<2> output_size,
                   ExpandingArray<2> kernel_size,
                   ExpandingArray<2> dilation,
                   ExpandingArray<2> padding,
                   ExpandingArray<2> stride) {
  if (input.dim() == 3) {
    return torch::col2im(
        input,
        output_size,

This file has been truncated. show original

InnovArul · July 11, 2021, 5:52am

Just incase if you want to dig deeper on col2im:

Native C++: pytorch/Col2Im.cpp at master · pytorch/pytorch · GitHub

Cuda code:
pytorch/Col2Im.cu at master · pytorch/pytorch · GitHub,
pytorch/im2col.cuh at master · pytorch/pytorch · GitHub

zfzhang · July 11, 2021, 11:16am

Thanks for the useful information!
A follow-up question is: If I want to write a modified version of col2im() with a different name, can I write it in C++ extensions? Or must I write it inside the ATen library and recompile?

InnovArul · July 11, 2021, 2:45pm

I haven’t tried it. In my understanding, it should be possible to write a modified version of col2im in a standalone way than inside ATen.

zfzhang · July 12, 2021, 4:46am

Good, thanks for the info!