C++ extension backward function

[Iaroslav_Elistratov]

I’m trying to create a cpp extension following this tutorial.

It seems that if we’re using torch::Tensor as input to a C++ function, and we’re only doing PyTorch operations on it, no custom backward function is necessary.

C++ backward function

if you are using torch::Tensor as input to your C++ function, and you are only doing PyTorch operations on it, no backward function is necessary (just like in Python you dont write one). However, if you are doing some custom calls, such as going into OpenCV and editing the contents, then custom backward call is necessary, as we dont know how to differentiate through those calls.

Yet the tutorial implements a backward pass even though, I believe, in the forward pass it uses torch::Tensor as input and uses only native PyTorch operations on it. So it seems we don’t need to implement the backward pass in this case but instead in lltm_backward we can just call .backward() on the inputs – is it correct?

[Munich]

@Iaroslav_Elistratov were you able to find the answer to this question? I have the exact same question after finding @smth response in the thread you posted above.