How to operate tensor from custom class in registered hook?


I implemented some subclass of pytorch tensor in my applications, and need to operate such custom tensors in a call back function hooked to a tensor’s backward pass. However, i found that any operation of such custom tensor would yield torch tensor inside such hooked callbacks, while i am hoping them to return tensor of the custom class.

A minimal example:

import torch

class CTensor(torch.Tensor):

def hook(grad):
    print('within hook:', type(a[:1]))

a = CTensor([1,2,3]).requires_grad_()
b = CTensor([1,2,3]).requires_grad_()

c = b.sum()
print('outside hook:', type(a[:1]))


This prints

outside hook: <class '__main__.CTensor'>
within hook: <class 'torch.Tensor'>

However, I would need the slicing operation inside the hook() also return tensor of my custom class CTensor. What should I do? Or this is something not possible?


@ptrblck Any help would be much appreciated!

I don’t know if PyTorch allows you to use custom tensors in Autograd.
You can extend the tensor class as described here, but I’m unsure if you could hook it somehow also into the backward pass. @albanD might know more about this use case and if it’s doable.

Yeah there is something fishy happening here, we should look into this in more details in the issue that is already opened with this example.