Will this inplace assignment propagate gradient correctly?

autograd
1

Hello,

In my code I need to do something like this

. . . 
# a (n_a, n_features)
# b (n_b, n_features)
# some_indices (n_a,) long

result = b.clone()
result[some_indices, :] = a
return result

Will this propagate gradients correctly for a and b ?
(ie the gradient of result for some_indices propagated to a, and for all the other indices, they are propagated to b)
I am a bit scared of the inplace operation, I couldn’t find explanations of where they’re safe or not for the autograd engine.

Usually I would use a torch.where, but a doesn’t have the same shape as b, so that would require me to get a to that same size, which would also require an assignment…

Thank you for the help

2

I would expect to see a RuntimeError if the inplace operation is disallowed and would result in wrong gradient calculations. If you are not seeing such an error in your code it should be fine.

3

Oh so if there is no error, I can expect the gradients to be correct? That is great!
Thanks!