Configs¶
Here is a tutorial on Google Colab that shows how to use the configs module
- class labml.configs.BaseConfigs(*, _primary: Optional[str] = None)[source]¶
You should sub-class this class to create your own configurations
- labml.configs.option(name: Union[any, List[any]])[source]¶
- labml.configs.option(name: Union[any, List[any]], option_name: str)
- labml.configs.option(name: Union[any, List[any]], pass_params: List[any])
- labml.configs.option(name: Union[any, List[any]], option_name: str, pass_params: List[any])
Use this as a decorator to register configuration options.
This has multiple overloads
- labml.configs.option(config_item: Union[any, List[any]])[source]
- labml.configs.option(config_item: Union[any, List[any]], option_name: str)[source]
- labml.configs.option(config_item: Union[any, List[any]], pass_params: List[any])[source]
- labml.configs.option(config_item: Union[any, List[any]], option_name: str, pass_params: List[any])[source]
- Parameters
name – the configuration item or a list of items. If it is a list of items the function should return tuple.
option_name (str, optional) – name of the option. If not provided it will be derived from the function name.
pass_params (list, optional) – list of params to be passed. If not provided the configs object is passed. If provided the corresponding calculated configuration items will be passed to the function
- labml.configs.calculate(name: Union[any, List[any]], func: Callable)[source]¶
- labml.configs.calculate(name: Union[any, List[any]], option_name: str, func: Callable)
- labml.configs.calculate(name: Union[any, List[any]], pass_params: List[any], func: Callable)
- labml.configs.calculate(name: Union[any, List[any]], option_name: str, pass_params: List[any], func: Callable)
Use this to register configuration options.
This has multiple overloads
- labml.configs.calculate(name: Union[any, List[any]], func: Callable)[source]
- labml.configs.calculate(name: Union[any, List[any]], option_name: str, func: Callable)[source]
- labml.configs.calculate(name: Union[any, List[any]], pass_params: List[any], func: Callable)[source]
- labml.configs.calculate(name: Union[any, List[any]], option_name: str, pass_params: List[any], func: Callable)[source]
- Parameters
name – the configuration item or a list of items. If it is a list of items the function should return tuple.
func – the function to calculate the configuration
option_name (str, optional) – name of the option. If not provided it will be derived from the function name.
pass_params (list, optional) – list of params to be passed. If not provided the configs object is passed. If provided the corresponding calculated configuration items will be passed to the function
- labml.configs.hyperparams(*args: any, is_hyperparam=True)[source]¶
Identifies configuration as (or not) hyper-parameters
- Parameters
*args – list of configurations
is_hyperparam (bool, optional) – whether the provided configuration items are hyper-parameters. Defaults to
True.
- labml.configs.meta_config(*args: any, is_meta=True)[source]¶
Identifies configuration as meta parameter
- Parameters
*args – list of configurations
is_meta (bool, optional) – whether the provided configuration items are meta. Defaults to
True.
- labml.configs.aggregate(name: any, option_name: str, *args: Tuple[any, any])[source]¶
Aggregate configs
- Parameters
name – name of the aggregate
option_name – aggregate option name
*args – list of configs to be aggregated
Dynamic Hyper-parameters¶
- class labml.configs.DynamicHyperParam(default: Union[float, int], type_: str, range_: Tuple[float, float])[source]¶
Dynamic hyper-parameters can be changed manually on labml.ai app <https://labml.ai> while the models are training. There are designed as an alternative to hyper-parameter schedules like gradually decreasing learning rate as the training progresses. Dynamic hyper-parameters let users change the values based on training metrics instead of sticking to a pre-determined schedule.
This is the base class that is extended by specific dynamic hyper-parameter types. You should use those.