Was the source code for F.conv2d() implemented by Linear transformation?

amauhaha · August 22, 2020, 10:45am

Hello,

we know that convolution can be regarded as the special case of Linear transformation.
I am not really sure whether the source code for F.conv2d() was implemented by Linear transformation ?

I consider the code (it looks like the source code for F.conv2d())

github.com

pytorch/pytorch/blob/master/aten/src/ATen/native/Convolution.cpp

#include <limits>
#include <ATen/ATen.h>
#include <ATen/NativeFunctions.h>
#include <ATen/native/cpu/DepthwiseConvKernel.h>
#include <ATen/native/utils/ParamUtils.h>
#include <ATen/native/ConvUtils.h>
#include <ATen/native/xnnpack/Engine.h>

#include <ATen/Config.h>
#include <c10/macros/Macros.h>

#if AT_NNPACK_ENABLED()
#include <nnpack.h>
#endif


constexpr int MIOPEN_DIM_MAX = 5;

namespace at { namespace native {

This file has been truncated. show original

github.com

pytorch/pytorch/blob/e9902358df14dc4809e4f50b12088a5200a1862d/aten/src/ATen/native/ConvolutionMM2d.cpp

#include <ATen/ATen.h>
#include <ATen/AccumulateType.h>
#include <ATen/Dispatch.h>
#include <ATen/Parallel.h>
#include <ATen/TensorUtils.h>
#include <ATen/core/grad_mode.h>
#include <ATen/div_rtn.h>
#include <ATen/native/Unfold2d.h>

namespace at {
namespace native {

namespace {

static inline void slow_conv2d_shape_check(
    const Tensor& input,
    const Tensor& grad_output,
    const Tensor& weight,
    const Tensor& bias,
    int64_t kernel_height,

This file has been truncated. show original

However, I still have no idea about it