plugins.base

Base class for all plugins.

Plugin Objects

class Plugin()

Base class for all plugins.

It is structured around three core components [trn.Trainer, module_interface.ModuleInterface, data_interface.DataInterface]. Derived classes should implement the methods setup_data, setup_module, and setup. These methods will execute the data processing pipeline and initialize the required components for training such as trainer and module_interface. setup_trainer initializes the PyMarlin trainer and backend.

plugin.setup is provided to bootstrap the entire pipeline for a specific downstream task. Example::

trainer = plugin.setup() trainer.train() trainer.validate()

__init__

def __init__(config: Optional[Dict] = None)

CustomArgParser parses YAML config located at cmdline --config_path. If --config_path is not provided, assumes YAML file is named config.yaml and present in working directory. self.trainer_args (trn.TrainerArguments): Instantiated dataclass containing args required to initialize trn.Trainer class.

datainterface

@property

def datainterface()

DataInterface object used for data processing. The property can be set in setup_datainterface.

Returns:

An object of type data_interface.DataInterface.

dataprocessor

@property

def dataprocessor()

DataProcessor object(s) used for data processing. The property may be used in conjuction with datainterface in the setup_datainterface method.

Returns:

An object of type data_interface.DataProcessor.

moduleinterface

@property

def moduleinterface()

ModuleInterface object. The property can be set in setup_module.

Returns:

An object of type module_interface.ModuleInterface.

trainer

@property

def trainer()

Trainer object. The property can be set in setup_trainer.

Returns:

An object of type trn.Trainer.

setup_datainterface

@abstractmethod

def setup_datainterface(*args: Optional)

Derived plugins must implement this method. The method should execute a generic data processing pipeline for the task and update the TaskDataInterface object to contain the processed train and val datasets.

NOTE to TaskPlugin designers: Typically, the plugin shouldn't need any input arguments from user except from the YAML config. DataInterface and DataProcessor related arguments should be processed in the init method of the TaskPlugin.

Returns:

• datainterface_obj data_interface.DataInterface - TaskDataInterface object

setup_module

@abstractmethod

def setup_module(*args: Optional)

Derived plugins must implement this method. The method should create a TaskModuleInterface object (module_interface.ModuleInterface) and set moduleinterface property.

NOTE to TaskPlugin designers: Typically, the plugin shouldn't need any input arguments from user. ModuleInterface related arguments should be processed in the init method of the TaskPlugin.

setup_trainer

def setup_trainer()

Creates a trn.Trainer object and sets the trainer property. Used by all plugins unless overriden (not recommended).

setup

@abstractmethod

def setup(**kwargs)

Executes all steps from data processing to trainer initialization.

This should be equivalent to::

plugin.setup_datainterface()

plugin.setup_module()

plugin.setup_trainer()