Is there any good documentation about handling torch::autograd::FunctionTask?

autograd
1

Is there any good documentation about torch::autograd::FunctionTask?

2

I don’t think there is any. This is an internal class.
Where did you encountered it? What do you want to do with it?

3

Thank you for commenting.
I try to implement a function for forward/backward.
I met following error message.

RuntimeError: Function TestBackward returned an invalid gradient at index 0 - expected shape [1000] but got [1, 1000]

Then I investigate it and found it seems InputBuffer in FunctionTask is not expected shape. (specifically metadata.shape() and grads[i].sizes() inconsistency.)
Is there any tips to fix InputBuffer tensor size?

4

The InputBuffer size will be the size of what the previous function returned.
Make sure that the gradient you return is exactly the same as the input it is for: 1x1000 is not the same as 1000.