xlm.utils.saving_utils

mkdir_rank_zero_only(dir, exist_ok=True)

Create directory only on rank 0.

Parameters:

Name	Type	Description	Default
dir	Path	Directory path.	required
exist_ok	bool	If True, do not raise an exception if the directory already exists. Default to True.	True

process_state_dict(state_dict, symbols=0, exceptions=None)

Filter and map model state dict keys.

Parameters:

Name	Type	Description	Default
state_dict	Union[OrderedDict, dict]	State dict.	required
symbols	int	Determines how many symbols should be cut in the beginning of state dict keys. Default to 0.	0
exceptions	Union[str, List[str]]	Determines exceptions, i.e. substrings, which keys should not contain.	None

Returns:

Name	Type	Description
OrderedDict	OrderedDict	Filtered state dict.

save_predictions_from_dataloader(predictions, path)

Save predictions returned by Trainer.predict method for single dataloader.

Parameters:

Name	Type	Description	Default
predictions	List[Any]	Predictions returned by Trainer.predict method.	required
path	Path	Path to predictions.	required

save_predictions(predictions, dirname, output_format='json')

Save predictions returned by Trainer.predict method.

Due to LightningDataModule.predict_dataloader return type is Union[DataLoader, List[DataLoader]], so Trainer.predict method can return a list of dictionaries, one for each provided batch containing their respective predictions, or a list of lists, one for each provided dataloader containing their respective predictions, where each list contains dictionaries.

Parameters:

Name	Type	Description	Default
predictions	List[Any]	Predictions returned by Trainer.predict method.	required
dirname	str	Dirname for predictions.	required
output_format	str	Output file format. It could be json or csv. Default to json.	'json'