cli

Functions

Classes

class pytorch_lightning.utilities.cli.LightningArgumentParser(*args, parse_as_dict=True, **kwargs)[source]

Bases: jsonargparse.

Extension of jsonargparse’s ArgumentParser for pytorch-lightning

Initialize argument parser that supports configuration file input

For full details of accepted arguments see ArgumentParser.__init__.

add_lightning_class_args(lightning_class, nested_key, subclass_mode=False)[source]

Adds arguments from a lightning class to a nested key of the parser

Parameters

• lightning_class (Union[Callable[..., Union[Trainer, LightningModule, LightningDataModule, Callback]], Type[Trainer], Type[LightningModule], Type[LightningDataModule], Type[Callback]]) – A callable or any subclass of {Trainer, LightningModule, LightningDataModule, Callback}.

• nested_key (str) – Name of the nested namespace to store arguments.

• subclass_mode (bool) – Whether allow any subclass of the given class.

Return type

List[str]

add_lr_scheduler_args(lr_scheduler_class, nested_key='lr_scheduler', link_to='AUTOMATIC')[source]

Adds arguments from a learning rate scheduler class to a nested key of the parser

Parameters

• lr_scheduler_class (Union[Type[_LRScheduler], Type[ReduceLROnPlateau], Tuple[Union[Type[_LRScheduler], Type[ReduceLROnPlateau]], ...]]) – Any subclass of torch.optim.lr_scheduler.{_LRScheduler, ReduceLROnPlateau}.

• nested_key (str) – Name of the nested namespace to store arguments.

• link_to (str) – Dot notation of a parser key to set arguments or AUTOMATIC.

Return type

None

add_optimizer_args(optimizer_class, nested_key='optimizer', link_to='AUTOMATIC')[source]

Adds arguments from an optimizer class to a nested key of the parser

Parameters

• optimizer_class (Union[Type[Optimizer], Tuple[Type[Optimizer], ...]]) – Any subclass of torch.optim.Optimizer.

• nested_key (str) – Name of the nested namespace to store arguments.

• link_to (str) – Dot notation of a parser key to set arguments or AUTOMATIC.

Return type

None

class pytorch_lightning.utilities.cli.LightningCLI(model_class, datamodule_class=None, save_config_callback=<class 'pytorch_lightning.utilities.cli.SaveConfigCallback'>, save_config_filename='config.yaml', save_config_overwrite=False, trainer_class=<class 'pytorch_lightning.trainer.trainer.Trainer'>, trainer_defaults=None, seed_everything_default=None, description='pytorch-lightning trainer command line tool', env_prefix='PL', env_parse=False, parser_kwargs=None, subclass_mode_model=False, subclass_mode_data=False)[source]

Bases: object

Implementation of a configurable command line tool for pytorch-lightning

Receives as input pytorch-lightning classes (or callables which return pytorch-lightning classes), which are called / instantiated using a parsed configuration file and / or command line args and then runs trainer.fit. Parsing of configuration from environment variables can be enabled by setting env_parse=True. A full configuration yaml would be parsed from PL_CONFIG if set. Individual settings are so parsed from variables named for example PL_TRAINER__MAX_EPOCHS.

Example, first implement the trainer.py tool as:

from mymodels import MyModel
from pytorch_lightning.utilities.cli import LightningCLI
LightningCLI(MyModel)

Then in a shell, run the tool with the desired configuration:

$ python trainer.py --print_config > config.yaml
$ nano config.yaml  # modify the config as desired
$ python trainer.py --cfg config.yaml

Warning

LightningCLI is in beta and subject to change.

Parameters

• model_class (Union[Type[LightningModule], Callable[..., LightningModule]]) – LightningModule class to train on or a callable which returns a LightningModule instance when called.

• datamodule_class (Union[Type[LightningDataModule], Callable[..., LightningDataModule], None]) – An optional LightningDataModule class or a callable which returns a LightningDataModule instance when called.

• save_config_callback (Optional[Type[SaveConfigCallback]]) – A callback class to save the training config.

• save_config_filename (str) – Filename for the config file.

• save_config_overwrite (bool) – Whether to overwrite an existing config file.

• trainer_class (Union[Type[Trainer], Callable[..., Trainer]]) – An optional subclass of the Trainer class or a callable which returns a Trainer instance when called.

• trainer_defaults (Optional[Dict[str, Any]]) – Set to override Trainer defaults or add persistent callbacks.

• seed_everything_default (Optional[int]) – Default value for the seed_everything() seed argument.

• description (str) – Description of the tool shown when running --help.

• env_prefix (str) – Prefix for environment variables.

• env_parse (bool) – Whether environment variable parsing is enabled.

• parser_kwargs (Optional[Dict[str, Any]]) – Additional arguments to instantiate LightningArgumentParser.

• subclass_mode_model (bool) – Whether model can be any subclass of the given class.

• subclass_mode_data (bool) –

  Whether datamodule can be any subclass of the given class.

add_arguments_to_parser(parser)[source]

Implement to add extra arguments to parser or link arguments

Parameters

parser (LightningArgumentParser) – The argument parser object to which arguments can be added

Return type

None

add_configure_optimizers_method_to_model()[source]

Adds to the model an automatically generated configure_optimizers method

If a single optimizer and optionally a scheduler argument groups are added to the parser as ‘AUTOMATIC’, then a configure_optimizers method is automatically implemented in the model class.

Return type

None

add_core_arguments_to_parser()[source]

Adds arguments from the core classes to the parser

Return type

None

after_fit()[source]

Implement to run some code after fit has finished

Return type

None

before_fit()[source]

Implement to run some code before fit is started

Return type

None

before_instantiate_classes()[source]

Implement to run some code before instantiating the classes

Return type

None

fit()[source]

Runs fit of the instantiated trainer class and prepared fit keyword arguments

Return type

None

init_parser()[source]

Method that instantiates the argument parser

Return type

None

instantiate_classes()[source]

Instantiates the classes using settings from self.config

Return type

None

instantiate_trainer()[source]

Instantiates the trainer using self.config_init[‘trainer’]

Return type

None

link_optimizers_and_lr_schedulers()[source]

Creates argument links for optimizers and lr_schedulers that specified a link_to

Return type

None

parse_arguments()[source]

Parses command line arguments and stores it in self.config

Return type

None

prepare_fit_kwargs()[source]

Prepares fit_kwargs including datamodule using self.config_init[‘data’] if given

Return type

None

class pytorch_lightning.utilities.cli.SaveConfigCallback(parser, config, config_filename, overwrite=False)[source]

Bases: pytorch_lightning.callbacks.base.Callback

Saves a LightningCLI config to the log_dir when training starts

Raises

RuntimeError – If the config file already exists in the directory to avoid overwriting a previous run

setup(trainer, pl_module, stage=None)[source]

Called when fit, validate, test, predict, or tune begins

Return type

None

pytorch_lightning.utilities.cli.instantiate_class(args, init)[source]

Instantiates a class with the given args and init.

Parameters

• args (Union[Any, Tuple[Any, ...]]) – Positional arguments required for instantiation.

• init (Dict[str, Any]) – Dict of the form {“class_path”:…,”init_args”:…}.

Return type

Any

Returns

The instantiated class object.