Update nested config of ingredient

Currently I work with experiments which inject multiple ingredients. Each of the ingredients has its own YAML config file which is loaded via the ingredient.add_config method. For example ingredient.yaml:

...
solver_cfg:
  ada_lr:
    factor: 0.1
    epochs: [] 
...

Now I want to change one of the nested configuration parameters only for one of the experiments. The change should also be in the respective experiment.yaml:

...
update_ingredient:
  solver_cfg:
    ada_lr:
      epochs: [100] 
...

Is there a way to achieve this with the currently available methods? config_hook and option_hook only work on config updates which are implied by the command line. I could put the updates in a separate YAML file and add this to the ingredient only for the experiment but this will overwrite and not update the solver_cfg entry which loses the factor parameter. Am I missing something?

[Qwlouse]

In the short-term ingredients will remain the same, because I currently don’t have the time for an in-depth refactoring. In the medium term, I would like to make ingredients classes, that you include by instantiating them as part of the configuration process. This would make their usage very flexible and hopefully intuitive. Something along these lines:

class MyDataset(Ingredient):
    def __init__(self, variant="variant1"):
        self.path = "foo/bar/{}".format(variant)
        self.normalize = True

@ex.config
def cfg():
    learn_rate = 0.1
    problem = "A"
    if problem == "A":
        dataset = MyDataset()
    elif problem == "B":
        dataset = MyDataset(variant="variant2")
    else:
        dataset = MyOtherDataset()

Which could result in a config like this:

Configuration:
    learn_rate = 0.1
    problem = "B"
    seed = 1234
    dataset:
        path = "foo/bar/variant2"
        normalize = True

[Qwlouse]

Closing this issue, since we sort-of solved your problem, and solving it completely probably requires larger changes that deserve their own issues.