Parameter Classes

Parameters are used to control every module in tfaip including hyper-paramters of training, the selection of graph, modifying the data pipeline, etc. Each module is split into a class providing the actual implementation, and a parameter dataclass that only holds its parameters, e.g. Trainer and TrainerParams. Each Params-class of tfaip is wrapped by dataclass and pai_dataclass which allows to easily add new parameters that support typing and will automatically be accessible from the command line. Command line support is provided by :paiargparse which automatically converts a dataclass hierarchy into command line arguments.

In the following, all primary parameter classes are listed

• ScenarioBaseParams * DataBaseParams * ModelParams * EvaluatorParams

• TrainerParams

  • DeviceParams

  • OptimizerParams

  • LearningRateParams

  • WarmstartParams

• LAVParams

• PredictorParams

Parameters

All parameters that shall be available during training or loading of a model (e.g., resume training, or LAV) must be added to a params class within the hierarchy (see above).

Custom Parameter Classes

You are allowed to add additional parameter classes which must wrap @pai_dataclass and @dataclass. Make sure to use a descriptive name for the name and the field storing that dataclass, e.g.:

Subclassing Parameter Classes

Subclassing a dataclass will inherit all fields of the parent. tfaip uses this to extend parameters for almost every module. Note, that this is the reason why every field must have at least a dummy default value. Required parameters to be set from the command line must be specified by setting the metadata: field(..., metadata=pai_meta(required=True)).

Hidden Parameters

To hide a parameter from the command line, set its field metadata mode to ignore: field(..., metadata=pai_meta(mode="ignore"). These parameters are considered to be state parameters that are set or changed during training (e.g., TrainerParams.export_best_state) or are set by the scenario to exchange information (e.g., the DataParams could define num_classes which are passed to ModelParams.num_classes during the initialization). To share parameters implement the __post_init__ function of a parent dataclass if possible, else override ScenarioBase.create_model, ScenarioBase.create_data, etc. to specify a parameter directly before the actual class is instantiated.

Command Line

The parameter hierarchy is parsed and flattened to allow to set the parameters from the command line. The following types are supported, see here for a full list and examples: * Primitive types: str, int, float, bool * Enums: IntEnum, StrEnum * Lists: List[str], List[int], List[float], List[Enum] * Other dataclasses defined with @pai_dataclass and @dataclass, also in Dicts and Lists

Naming convention: * Dataclasses in snake mode, the default (dc_meta(mode="snake")) are added as snake mode, e.g. --train.batch_size * Dataclasses in flat mode (pai_meta(mode="flat")) are added as root parameter, e.g. --model or --trainer

Meta data can be specified to for example add a help string or change the mode.

Example: