How to do a basic config/parameter validation

[hogru]

I found a couple of ways to validate my configuration parameters (Enums, custom OmegaConf resolvers, OmegaConf.to_object()). However I fail to do even a simple type check. I would expect this to raise an exception or at least give me type hints in the IDE, but apparently I again don't see the forest for the trees. Can you help me in telling me what's wrong with this code?

from dataclasses import dataclass

import hydra
from omegaconf import OmegaConf


@dataclass
class MyParam:
    my_param: int  # I want my_param to be an int


@hydra.main(config_path="./", config_name="test", version_base=None)
def main(cfg: MyParam) -> None:
    print(OmegaConf.to_yaml(cfg))  # my_param: this_is_a_string
    print(type(cfg.my_param))      # <class 'str'>


if __name__ == "__main__":
    main()

And this is test.yaml

my_param: this_is_a_string

[Jasha10]

See the docs on Structured Config schema.
You'll need to use a schema if you want to take advantage of OmegaConf's runtime type checking.

[hogru]

I thought that I had a schema… until it did not catch type errors. The consequence of this problem was my attempt to provide a minimum “working” example above and “loosing” the schema during the process. Your hint encouraged me to take another look at the structured config schema, and I realized that I had missed the correct cs.store() calls along with the corresponding defaults entries. I must admit this is one of the more complex frameworks I came across, I think I am only starting to fully understand what's going on behind the scenes (schema, multiple .yaml and .py files, …). But now I get ValidationError exceptions, which is what I wanted. Thank you! Next: __post_init__() ;-)

[Jasha10]

I must admit this is one of the more complex frameworks I came across

Hydra is a framework for building frameworks ;)

Note you'll need to use OmegaConf.to_object if you want MyParam.__post_init__ to take effect. OmegaConf.to_object will convert your config from an instance of DictConfig to an instance of MyParams. For example:

...

@dataclass
class MyParam:
    ...

    def __post_init__(self):
        ...

cs.store(name="my_param_schema", node=MyParam)

@hydra.main(...)
def main(_cfg: DictConfig):
    assert isinstance(_cfg, DictConfig)
    cfg: MyParams = OmegaConf.to_object(_cfg)  # __post_init__ runs here
    ...
    

[hogru]

Thanks, @Jasha10, your “meta comment” made me feel a bit more capable 😉

You already "told" me about OmegaConf.to_object() in this discussion. This works nicely and I use it to do some validations I can't (or don't know how to) do within the dataclasses, e.g. assert sum(random_split) = 1.. But for actually accessing the config values, I keep the structured config because I feel its syntax is nicer to use. So, the only thing that remains is that I would like to convert some config values to e.g., NamedTuples. This is not strictly necessary, but just easier to read. I do this in main() now because doing it in __post_init__() works on the data structure that I don't use to access the config values (for the reason above). Thanks again for your help 👍