RankCallback

class lightning_ir.callbacks.callbacks.RankCallback(save_dir: Path | str | None = None, run_name: str | None = None, overwrite: bool = False)[source]

Bases: Callback, _GatherMixin, _OverwriteMixin

__init__(save_dir: Path | str | None = None, run_name: str | None = None, overwrite: bool = False) → None[source]

Callback to write run file of ranked documents to disk.

Parameters:

  • save_dir (Path | str | None) – Directory to save run files to. If None, run files will be saved in the models’ directory. Defaults to None.

  • run_name (str | None) – Name of the run file. If None, the dataset’s dataset_id or file name will be used. Defaults to None.

  • overwrite (bool) – Whether to skip or overwrite already existing run files. Defaults to False.

Methods

__init__([save_dir, run_name, overwrite])

Callback to write run file of ranked documents to disk.

on_test_batch_end(trainer, pl_module, ...[, ...])

Hook to aggregate and write ranking to file.

setup(trainer, pl_module, stage)

Hook to setup the callback.

teardown(trainer, pl_module, stage)

Hook to cleanup the callback.

Attributes

on_test_batch_end(trainer: Trainer, pl_module: LightningIRModule, outputs: LightningIROutput, batch: Any, batch_idx: int, dataloader_idx: int = 0) → None[source]

Hook to aggregate and write ranking to file.

Parameters:

  • trainer (Trainer) – PyTorch Lightning Trainer.

  • pl_module (LightningIRModule) – LightningIR module.

  • outputs (LightningIROutput) – Scored query documents pairs.

  • batch (Any) – Batch of input data.

  • batch_idx (int) – Index of batch in the current dataset.

  • dataloader_idx (int | None) – Index of the dataloader. Defaults to 0.

Raises:

ValueError – If the batch does not have query_ids.

setup(trainer: Trainer, pl_module: LightningIRModule, stage: str) → None[source]

Hook to setup the callback.

Parameters:

  • trainer (Trainer) – PyTorch Lightning Trainer.

  • pl_module (LightningIRModule) – LightningIR module.

  • stage (str) – Stage of the trainer, must be “test”.

Raises:

  • ValueError – If the stage is not “test”.

  • ValueError – If no save_dir is provided and model_name_or_path is not a path (the model is not local).

teardown(trainer: Trainer, pl_module: BiEncoderModule, stage: str) → None[source]

Hook to cleanup the callback.

Parameters:

  • trainer (Trainer) – PyTorch Lightning Trainer.

  • pl_module (LightningIRModule) – LightningIR bi-encoder module used for indexing.

  • stage (str) – Stage of the trainer, must be “test”.