Dataconfigs

About

Dataconfigs provides a seamless integration between dataclasses as configs and any configurable classes that use these configs. The most powerful feature is that config attributes become part of configurable, i.e., no need for explicit .config attribute (unless specified otherwise). This keep class definitions very siple yet decouples configurable parameters from class attributes. In terms of configs, it is a simplified hydra alternative, yet supporting more features.

Features

Last 3 features are in development:

• Embedded config params in configurables (no need for separate .config attribute, unless specified otherwise)
• Easy type annotations (visible annotations, even though configs are never inherited from)
• Support for multiple, nested, hierarchical, and union configs, as well as non-atomic types
• Various file support (json, yaml, pickle)
• CLI support
• Decorator-based (inspired by Hydra)

Quick Start

Install the library:

pip install dataconfigs

Check the demo.

Simple Script Example

1. Simply define your config (must be a datacalss and contain the word Config in name!)
2. Extend the config with your configurable (it won't turn into a dataclass unless specified!)
3. Instantiate your configurable by optionally overwriting default config parameters

from datacalsses import dataclass, is_dataclass
from dataconfigs import configurable, show_config_params

@datacalss
class MyConfig:
    """My config class
    Args:
        param1 (int): The first parameter.
        param2 (str): The second parameter.
    """
    param1: int = 0
    param2: str = "param2"

@configurable
class MyConfigurable(MyConfig):
    def __init__(self, attr1=1, attr2=2):
        print(self.param1, self.param2) # Already available!

assert not is_dataclass(MyConfigurable) # Only attributes were copied!
obj = MyConfigurable("attr1", param1=1) # Can overwrite config params! 
# 1, 'param2'

Simple CLI Example (only ideas for now)

Assume you have config.json file as follows:

{
    "param1": 111,
    "param2": "222"
}

One possibility is to directly make the configurable. Example main.py:

@configurable.main("path/to/config.json", MyConfigurable, attr1="attr1")
def main(my_configurable):
    pass

if __name__ == "__main__":
    main()

We can still overwrite config params or the whole config file when calling:

dataconfigs main.py --param2 'custom' # 111, 'custom'
dataconfigs main.py --config path/to/config2.yaml

If we call dataconfigs main.py -h, descriptions will be extracted from docstrings:

Options:
--param1 : (int) The first parameter. Defaults to 111.
--param2 : (str) The second parameter. Defaults to '222'.

Tutorials (TODO)

1. Config classes: there are many ways to define configs. The only restriction is that they must be dataclasses and contain an "Config" word in their class name (see basics):
   • Basics
   • Unions & non-atomic types
   • Multiple, nested and hierarchical configs
   • Config files (json, yaml, pkl, etc.)
2. Configurables: any class (configurable) that extends the config(-s) will only obtain its attributes and won't alter its class definition. The configurable class can inherit from any other parents in the usual manner.
   • Config vs non-config attributes
   • Various ways to instantiate
3. Compatibility with Third Party CLI
   • Pytorch Lightning
   • Tensorflow
   • Optuna
   • Ray