[Python] How to avoid redundancy when using Hydra structured configs with schema validation?

I'm using Hydra in my Python project and started adding schema validation using structured configs (i.e., defining config classes with @dataclass and type annotations).

While this adds type safety and catches config errors early, I'm finding it quite redundant. Every time I update something in my .yaml config files (e.g., renaming a field or adding a parameter), I have to go back and make the exact same change in the corresponding structured config class. It feels like double the work.

Example:

model:
hidden_size: 128
num_layers: 3


And then I have to mirror this in Python:


@dataclass
class ModelConfig:
hidden_size: int = 128
num_layers: int = 3


Is this just the cost of having validation/type safety with Hydra? Or is there a better workflow or tooling that can reduce the duplication between YAML and the Python schema?

How do people handle this in larger projects where configs change often?

Continue reading...