Writing a small Pytorch extension in C++

C++
1

Hi

I’m trying to write a small extension to PyTorch in C++. I’ve tried to look at the examples provided and see that the syntax is quite similar to what the python syntax is.

Part - 1
Here is my code for the extension

#include <torch/extension.h>
#include <vector>
#include <iostream>


torch::Tensor compute_grad(torch::Tensor in, torch::Tensor out)
{
    std::cout << "Input is:" << in << "\n" << "Output is: " << out << "\n";
    auto b = out.sum();
    std::cout << "Out_sum:" << "\t" << b << std::endl;
    b.backward();
    std::cout << __LINE__ << std::endl;
    return in.grad();
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
{
    m.def("compute_grad", &compute_grad, "Compute Gradient");
}

and I call it with

import torch
import DenseGrad

if __name__ == "__main__":
    x = (2 * torch.ones(2, 2)).requires_grad_()
    y = x ** 3
    print(DenseGrad.compute_grad(x, y))

For some reason, the backward() call is getting stuck and the only way to kill it is through top. Can someone tell me what I’m doing wrong?

2

I tried running your program with gdb, I’m getting a deadlock on the autograd execution engine. There are some posts related to this one, and a currently opened issue on PyTorch’s GitHub.

It seems to me that people have solved this by changing the environment, it could be related to PyTorch’s build and dependencies.

3

Hi!
Thanks for taking the time to put it through gdb. From this, it seems to have been fixed a while ago, and I’m using conda to get the latest version of pytorch (1.1.0).

I’m unsure how to go about it though.

4

Hi there. @Haydnspass and me have created a simple example extension. You can find it here. Hope it helps