Passing tuple/int to custom conv forward

C++
1

Hi,
I’m trying to use the C++ Extension tutorial to write a custom convolution function for use in python.

torch::Tensor conv2d_forward(
    const torch::Tensor input,
    const torch::Tensor weight,
    const torch::Tensor bias,
    torch::ExpandingArray<2> stride,
    torch::ExpandingArray<2> padding,
    torch::ExpandingArray<2> dilation,
    int64_t groups
){
    return F::conv2d(input,weight,F::Conv2dFuncOptions().bias(bias).stride(stride).padding(padding).dilation(dilation).groups(groups));
}

I was hoping that I could then call this function using pybind like this:

torchexplain_cpp.conv2d_forward(input, weight, bias, (1,1),(1,1),(1,1),1)

But I get an incompatible arguments error regarding the tuples.
How would I appropriately pass python tuples OR ints to the function?
Thank you

2

Could you try to use c10::ArrayRef<long int> as the type for stride, padding, and dilation and rerun the script?

3

Thank you for your reply, that works now. I’m still new to C++ and I realise now that ExpandingArray is a template. Is it possible to use it for parameter types? or would I have to either:
Convert the ints to tuples in python
or
Write separate functions in C++ for int options vs. tuple options?

4

If you would like to pass integers directly, you could use the int64_t type in C++, as you’ve already done with the groups parameter.
Could you explain the use case a bit, if I misunderstood it?

5

Of course sorry. To clarify I was hoping that by specifying the optional parameters as ExpandingArray, then the function would accept either tuples, or integers that would then be expanded to tuples. So I could either pass (1,1) or simply 1 for my stride and see the same behaviour.
Again I’m still learning c++ and it seems now incorrect.
Thank you again for your help!

6

Ah OK, thanks for the explanation.
I’m not sure what the proper way would be, but I’ve seen the usage of c10::ArrayRef<long int> in C++ and passing a tuple with different number of elements on the Python side via:

(1,)
(1, 2)
(1, 2, 3)
...