Can bool slices operation be placed inside a back propagation loop


I just encountered an error, when I perform a bool slices operation like:

average = torch.mean(tensor[tensor[:] != 0]),

the system returns me an error: Function ‘AddmmBackward’ returned nan values in its 2th output.
I realized that this equation may not be a differentiable function, so I was wondering if there is a differentiable way can let me get the same result as the equation stated above.



This operation is fine. Note that you don’t need the [:].
This error occurs when you run in AnomalyMode? You should check which forward function it correspond to and print the output value to make sure they are in acceptable ranges.


Thanks for your suggestions and this operation really worked after I fixed a little error somewhere else.