hydra_train

allennlp_hydra.commands.hydra_train

The hydra-train creates an AllenNLP config by composing a set of yaml files with Hydra's Compose API then passing it to AllenNLP's train method.

Overriding using Hydra's Override Grammar is also supported.

The hydra-train command uses the underlying functionality of the compose command. So the use that documentation for understanding the composition of .yaml files.

Parameters¶

config_path : Union[str, PathLike]
Path to the root config directory.
config_name : str
The name of the root config file. Do NOT include the .yaml.
job_name : str
The job name. This is passed to Hydra and is not used here.
-s/--serialization-dir : Union[str, PathLike]
The directory where everything is saved.
-r/--recover : bool, optional (default = False)
Flag. Recover training from the state in serialization_dir
-f/--force : bool, optional (default = False)
Flag. Overwrite the output directory if it exists.
--node-rank : int, optional (default = 0)
The rank of this node in the distributed setup
--dry-run : bool, optional (default = False)
Flag. Do not train the model, but create a vocabulary, show dataset statistics and other training information
--file-friendly-logging : bool, optional (default = False)
Flag. Outputs tqdm status on separate lines and slows tqdm refresh rate
-o/--overrides : List[str], optional (default = [])
Keyword arguments passed will be used as a list of overrides using Hydra's override grammar for the config.

Example usage:
```
--overrides A=B C="D"
```
Will be interpreted as overrides ['A=B', 'C="D"']

HydraTrain#

@Subcommand.register("hydra-train")
class HydraTrain(Subcommand)

add_subparser#

class HydraTrain(Subcommand):
 | ...
 | @overrides
 | def add_subparser(
 |     self,
 |     parser: argparse._SubParsersAction
 | ) -> argparse.ArgumentParser

hydra_train_model_from_args#

def hydra_train_model_from_args(args: argparse.Namespace)

Just converts from an argparse.Namespace object to string paths.