Overriding values of non-primitive type in experiment

[DirkKuhn]

Hi,

I am trying to override the activation function of my neural network in an experiment group. Unfortunately, I seem to always run into the error

hydra.errors.ConfigCompositionException: In 'experiment/selu': ReadonlyConfigError raised while composing config:
Cannot set value of read-only config node
    full_key: model.activation_fn.path
    object_type=Just

Here is the corresponding minimal example.

from typing import Type
from torch import nn

from hydra_zen import to_yaml, store, builds, zen, launch, make_config
from omegaconf import DictConfig

class Model:
    def __init__(self, activation_fn: Type[nn.Module] = nn.ReLU):
        self.activation_fn = activation_fn

def app(zen_cfg: DictConfig, model: Model) -> None:
    print(to_yaml(zen_cfg, resolve=True))

store(Model, group='model')

Config = builds(
    app,
    populate_full_signature=True,
    hydra_defaults=[
        '_self_',
        {'model': 'Model'},
        {'experiment': 'selu'},
    ]
)

experiment_store = store(group='experiment', package='_global_')
experiment_store(
    make_config(
        hydra_defaults=['_self_'],
        model=dict(activation_fn=nn.SELU),
        bases=(Config,)
    ),
    name='selu'
)

if __name__ == '__main__':
    store.add_to_hydra_store()
    launch(Config, zen(app), version_base='1.3')

Is there anything I can do? I have already tried specifying the activation function in a separate config group. Is there maybe a way to make the node writable? Or do I have to pass the activation function as a partially instantiated object?
Thank you!

[rsokl]

Ah this maybe an unintentional side effect of just returning a frozen dataclass:

hydra-zen/src/hydra_zen/structured_configs/_implementations.py

Line 506 in 796d8ac

@dataclass(frozen=True)

If that is the case, I will happily change it to unblock this workflow.

[rsokl]

Nope! Just verifying before I move forward with making the fix in the library. Thanks!

[DirkKuhn]

Do you think you can make this modification the next couple of days? Or does it cause some issues?
Can you think of any disadvantages of passing the type instead of using partial configuration?
Again thanks so much!

[rsokl]

This should not cause any issues. I will try to get this in the latest pre-release tonight I hope.

[rsokl]

@DirkKuhn v0.11.0rc2 has been released and contains the fix to this issue. Thanks for your patience.

[DirkKuhn]

No worries. Thank you for your hard work :)