Hydra: Resetting the state of GlobalHydra

[chrdiller]

Hi,

I am trying to use the pytorch bottleneck profiler (python -m torch.utils.bottleneck train.py) which apparently has to run the hydra-decorated main() function twice and results in the error
AssertionError: GlobalHydra is already initialized

Is there a way to reset the state of the GlobalHydra singleton manually or another workaround to make the profiler work with hydra?

Thanks

[omry]

You can try calling the clear function at the end of your hydra.main function:

[chrdiller]

Ah cool. Had some trouble finding how to actually call this; in the end, hydra._internal.hydra.GlobalHydra.get_state().clear() worked.

[omry]

Yup.
Note that GlobalHydra is being promoted out of _internal in the next version and will be here.

[omry]

Ah, you are using it wrong (I suspect you are clearing all signletons which may have incidentally been what you want but not what you asked for).

try, in 0.11, try:
hydra._internal.hydra.GlobalHydra().clear()

In the next version it will be (I think):
hydra.core.global_hydra.GlobalHydra.instance().clear()

[chrdiller]

Okay, I also tried those two (I'm on version 0.11.3):

• hydra._internal.hydra.GlobalHydra().clear() gave me TypeError: clear() missing 1 required positional argument: 'self' (which makes sense I guess) and
• hydra._internal.hydra.GlobalHydra.instance().clear() raises AttributeError: type object 'GlobalHydra' has no attribute 'instance'

[chrdiller]

Ah sorry, hydra._internal.hydra.GlobalHydra.clear() doesn't work, hydra._internal.hydra.GlobalHydra().clear() does.

Thanks again!

[omry]

The above PR will fix this the next major version of Hydra.