How to validate my yaml file using hydra shema?

[HydrogenSulfate]

I prefer to use hydra with omegaconf cuz it's convenient to run multiple experiments. And I now want to validate the arguments in given yaml file by checking them and raise error if not correct or of wrong type.
For certain reason there are some arguments, except fixed args, not need to be checked(called "free args"), because name and number of "free args" may varies in different yaml file.

So how can I validate the fixed args and do not validate the free args using hydra + omegaconf with dataclass as schema? Suppose my yaml file is below:

mode: train # fixed arg, need to be validate
seed: 42  # fixed arg, need to be validate

scope:
   a: 1 # free arg, no need to be validate
   b: 2 # fixed arg, need to be validate

If this can not be achieved, is there any other method better than writting check function for each fixed arg?
Can anyone give an example code? Thanks a lot.

[Jasha10]

@HydrogenSulfate please read the docs on structured configs, which let you define a schema for validation.

• https://omegaconf.readthedocs.io/en/2.3_branch/structured_config.html
• https://hydra.cc/docs/tutorials/structured_config/intro/

[HydrogenSulfate]

Thanks for your reply, however I find it is still hard to make a "white_list" in which keys will be checked and otherwise ignored after I read these 2 tutorial. It might be a feature that hydra/omegaconf has not been implemented yet.

[HydrogenSulfate]

I guess I have found a solution: using pydantic BaseModel and custom validators for each item to be checked, and seems works well.

[odelalleau]

Structured configs should work if you can isolate the ignored keys under some section of your config, e.g. if in your example everything in scope could be ignored. But I don't think you can say "just validate these specific keys and ignore everything else" out of the box.

[HydrogenSulfate]

Code below will only check 'epochs' for given ConfigDict, and won't check any items which not defined in
schema Config.

class Config(BaseModel):
    epochs: Optional[int] = None

    @field_validator("epochs")
    def epochs_check(cls, v):
        # check 'epochs' via 'v' here ...
        return v

# pass DictConfig to pydantic schema
_cfg_pydantic = Config(**dict(config))

# convert back after checking passed
full_cfg = DictConfig(_cfg_pydantic.model_dump())

[HydrogenSulfate]

The key difference between pydantic and omegaconf is that omegaconf do not have(not default at least) a extra argument as in pydantic, which control whether allow config received in schema class(BaseModel or dataclass) can have extra fields.

[HydrogenSulfate]

@HydrogenSulfate please read the docs on structured configs, which let you define a schema for validation.

• https://omegaconf.readthedocs.io/en/2.3_branch/structured_config.html
• https://hydra.cc/docs/tutorials/structured_config/intro/