Slightly changing model causes 'modified by inplace operation' error

1

Previous code:
return n_vals

I added one line:
n_vals[:,1,:] = n_vals[:,1,:] * torch.tensor(2)

And now get:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

I don’t understand this error because the line I added is not in place.

Another potential cause of the error is because I am preloading the model’s weights from disk and changing the model (with the above line) after preloading the weights. Still, doesnt seem like it should cause this error…

2

The answer is that assigning the variable with slice notation is an inplace operation. The solution is to add a clone call:
n_vals[:,1,:] = n_vals[:,1,:].clone() * torch.tensor(2)