CODEWITHSUNDEEP
pythonfb-hydraomegaconf
Lukas
Lukas

Reputation: 543

Is it possible to use the override syntax with structured configs in Hydra?

The Hydra docs on configuring experiments explains how to use overrides in standard, file-base configs. Is it possible to use overrides in Structured Configs? A minimum example:

from dataclasses import dataclass
from hydra.core.config_store import ConfigStore
from omegaconf import DictConfig, MISSING

@dataclass
class MyConfig(DictConfig):
    data: DataConfig = MISSING
    model: ModelConfig = MISSING
    parameters: dict[str, Any] = MISSING

cs = ConfigStore.instance()

cs.store(name="my_conf", node=MyConfig)
cs.store(group="model", name="v1", node=ModelConfig_V1)
cs.store(group="data", name="synthetic", node=DataConfig_Synthetic)

I would like to add in experiments somewhat like this:

@dataclass
class ExperimentConfig_Default:
    description: str = "Some experiment."
    parameters: dict[str, Any] = field(default_factors=lambda:{"lr":0.01})

cs.store(group="experiment", name="default", node=ExperimentConfig_Default)

I want to be able to use the CLI like this:

python main.py model=v1 data=synthetic +experiment=default

to get a config like this:

data:
  ...
model:
  ...
parameters:
  lr: 0.01

Upvotes: 0

Views: 62

Answers (1)

Omry Yadan
Omry Yadan

Reputation: 33646

Yes, this is possible. At runtime the only difference between Structured Configs and file-based configs is that Structured Configs have runtime type checks for config composition/modification.

When registering a config in the config store, you can also specify the package to be global:

cs.store(
  group="experiment", 
  name="default", 
  node=ExperimentConfig_Default, 
  package="_global_"
)

Upvotes: 1

Related Questions