xlm.log_predictions

FilePredictionWriter

Bases: _PredictionWriter

Writer that outputs predictions to JSONL files.

__init__(fields_to_keep_in_output=None, file_path_='from_pl_module')

Initialize FilePredictionWriter.

Parameters:

Name	Type	Description	Default
fields_to_keep_in_output	Optional[List[str]]	List of fields to keep in the output. If None, all fields are kept.	None
file_path_	Union[Path, Literal['none', 'from_pl_module']]	Path to the file or special values. if "from_pl_module", query the pl_module for the predictions_file for the step and epoch set to "none" to disable file writing	'from_pl_module'

write(predictions, ground_truth_text, step, epoch, split, dataloader_name, pl_module, trainer)

Write predictions to a JSONL file.

Parameters:

Name	Type	Description	Default
predictions	List[Dict[str, Any]]	List of prediction dictionaries.	required
ground_truth_text	Optional[List[str]]	List of ground truth text strings.	required
step	int	Current training step.	required
epoch	int	Current epoch.	required
split	Literal['train', 'val', 'test', 'predict']	The split name (train/val/test/predict).	required
dataloader_name	str	The dataloader name.	required
pl_module	LightningModule	The Lightning module.	required
trainer	Optional[Trainer]	The Lightning trainer.	required

read(step, epoch, split, dataloader_name, pl_module)

Read predictions from a JSONL file.

LoggerPredictionWriter

Bases: _PredictionWriter

Writer that outputs predictions to Lightning loggers.

__init__(n_rows=10, logger_=None, fields_to_keep_in_output=None)

Initialize LoggerPredictionWriter.

Parameters:

Name	Type	Description	Default
n_rows	int	Number of rows to log to the logger.	10
logger_	Optional[List[Logger]]	List of loggers to use. If None, uses pl_module.trainer.loggers.	None
fields_to_keep_in_output	Optional[List[str]]	List of fields to keep in the output. If None, all fields are kept.	None

write(predictions, ground_truth_text, step, epoch, split, dataloader_name, pl_module, trainer)

Write predictions to Lightning loggers.

Parameters:

Name	Type	Description	Default
predictions	List[Dict[str, Any]]	List of prediction dictionaries.	required
ground_truth_text	List[str]	List of ground truth text strings.	required
step	int	Current training step.	required
epoch	int	Current epoch.	required
split	Literal['train', 'val', 'test', 'predict']	The split name (train/val/test/predict).	required
dataloader_name	str	The dataloader name.	required
pl_module	LightningModule	The Lightning module.	required
trainer	Optional[Trainer]	The Lightning trainer.	required

ConsolePredictionWriter

Bases: _PredictionWriter

Writer that outputs predictions to console.

__init__(fields_to_keep_in_output=None)

Initialize ConsolePredictionWriter.

Parameters:

Name	Type	Description	Default
fields_to_keep_in_output	Optional[List[str]]	List of fields to keep in the output. If None, all fields are kept.	None

write(predictions, ground_truth_text, step, epoch, split, dataloader_name, pl_module, trainer)

Write predictions to console.

Parameters:

Name	Type	Description	Default
predictions	List[Dict[str, Any]]	List of prediction dictionaries.	required
ground_truth_text	List[str]	List of ground truth text strings.	required
step	int	Current training step.	required
epoch	int	Current epoch.	required
split	Literal['train', 'val', 'test', 'predict']	The split name (train/val/test/predict).	required
dataloader_name	str	The dataloader name.	required
pl_module	LightningModule	The Lightning module.	required
trainer	Optional[Trainer]	The Lightning trainer.	required

LogPredictions

Main logging class that handles the shared pipeline and delegates to writers.

__init__(writers=None, inject_target=None, additional_fields_from_batch=None, fields_to_keep_in_output=None)

Initialize LogPredictions.

Parameters:

Name	Type	Description	Default
writers	Optional[Union[List[_PredictionWriter], List[Literal['file', 'logger', 'console']]]]	List of prediction writers. If None, creates default writers for backward compatibility.	None
inject_target	Optional[str]	Key in batch to use as ground truth. If None, empty strings are used.	None

__call__(pl_module, trainer, batch, preds, split, dataloader_name)

Log predictions using the shared pipeline and delegate to writers.

Parameters:

Name	Type	Description	Default
pl_module	LightningModule	The Lightning module.	required
trainer	Optional[Trainer]	The Lightning trainer.	required
batch	Dict[str, Any]	The input batch.	required
preds	Dict[str, Any]	The predictions.	required
split	Literal['train', 'val', 'test', 'predict']	The split name (train/val/test/predict).	required
dataloader_name	str	The dataloader name.	required

read(step, epoch, split, dataloader_name, pl_module)

Read predictions from the writers.