How to implement customized convolution operation

1

I am implementing a convolutional layer with customized convolution operator (such as deformable convolution). I find that I have to extend Pytorch by C and now I want to find the source code of convolution. Or is there any other simpler ways to do it?

Thank you!

1 Like
2

I have also tried to do so. Wait for better suggestions.

3

Waiting, too + 1 :blush:

4

You could try to use unfold/fold for a custom convolution. Note that it will be most likely slower than other optimized methods.

5

I would like to implement a custom convolution layer, where in I actually modify each individual part of the input with other external filter which also requires gradient.

I am succesful in implementing that in python, but it is very slow as expected.

Now to speed it up, I would like to implement the convolution in C++ front end. As shown here https://pytorch.org/tutorials/advanced/cpp_extension.html

But want to use the backward pass of the inbuilt convolution. Also the extra parameter which is a part of the layer should get the gradients as well.

Before starting it, I wanted to know whether I need to write my own backward pass, or does the c++ end take care of it.

Or is there any other simpler way, please suggest.
Thank you

6

There is a great reference for understanding Different kinds of Convolution Operators: