Here, we use tricks like a custom __new__ function to turn regular classes into config classes for hydra.
We then use instantiate() with _convert_="object" to instantiate these classes.
Except for the instantiate() call, most interesting things happen in model.py.
When using _convert_="object", a few rules have to be followed
(but there is an open issue to make these rules unnecessary).
First, the top level configuration object may not be marked for instantiation
(that is, it may not have a _target_ attribute).
Second, classes that are to be instantiated may not take structured configs (i.e., dataclasses or attrs classes)
as arguments.
Requires Python 3.10+ and hydra-core 1.3+.
The simplest valid invocation is this:
python run.py seed=0 data=cmnist
This uses the config values stored in conf/data/celeba/gender.yaml:
python run.py seed=1 data=celeba/gender
And this one uses conf/trainer/4gpus.yaml:
python run.py trainer=4gpus data=cmnist seed=0