Why are there two autograds in 0.12.1?

There seems to be 2 different autograds in mxnet 0.12.0. I couldnt really find it in the change log for 0.12.0 or 0.12.1:

from mxnet import autograd #has autograd.record()
from mxnet.contrib import autograd

Its slightly hard to tell what each one really does… In the scientific python community, we take NumPy’s lead in documenting the init file to describe what each module does. Maybe this is something MXNet can take up?

Hi dmadeka,
The correct one is "from mxnet import autograd"
The contrib folder is mainly for backward compatibility and experiments. I agree this could lead to some confusion and there should be a readme or as you suggested an init file detailing this information.

1 Like