InBatchCrossEntropy

class lightning_ir.loss.in_batch.InBatchCrossEntropy(pos_sampling_technique: Literal['all', 'first'] = 'all', neg_sampling_technique: Literal['all', 'first', 'all_and_non_first'] = 'all', max_num_neg_samples: int | None = None)[source]

Bases: InBatchLossFunction

In-batch cross-entropy loss function for ranking tasks. Originally proposed in: Fast Single-Class Classification and the Principle of Logit Separation

__init__(pos_sampling_technique: Literal['all', 'first'] = 'all', neg_sampling_technique: Literal['all', 'first', 'all_and_non_first'] = 'all', max_num_neg_samples: int | None = None)

Initialize the InBatchLossFunction.

Parameters:

  • pos_sampling_technique (Literal["all", "first"]) – Technique for positive sample sampling.

  • neg_sampling_technique (Literal["all", "first", "all_and_non_first"]) – Technique for negative sample sampling.

  • max_num_neg_samples (int | None) – Maximum number of negative samples to consider. If None, all negative samples are considered.

Raises:

ValueError – If the negative sampling technique is invalid for the given positive sampling technique.

Methods

compute_loss(output)

Compute the in-batch cross-entropy loss.
compute_loss(output: LightningIROutput) → torch.Tensor[source]

Compute the in-batch cross-entropy loss.

Parameters:

output (LightningIROutput) – The output from the model containing scores.

Returns:

The computed loss.

Return type:

torch.Tensor

get_ib_idcs(output: LightningIROutput, batch: TrainBatch) → Tuple[torch.Tensor, torch.Tensor]

Get in-batch indices for positive and negative samples.

Parameters:

  • output (LightningIROutput) – The output from the model containing scores.

  • batch (TrainBatch) – The training batch containing targets.

Returns:

Indices of positive and negative samples.

Return type:

Tuple[torch.Tensor, torch.Tensor]

Raises:

ValueError – If scores are not present in the output.

process_scores(output: LightningIROutput) → torch.Tensor

Process the scores from the output.

Parameters:

output (LightningIROutput) – The output from the model.

Returns:

The scores tensor.

Return type:

torch.Tensor

process_targets(scores: torch.Tensor, batch: TrainBatch) → torch.Tensor

Process the targets from the batch.

Parameters:

  • scores (torch.Tensor) – The scores tensor.

  • batch (TrainBatch) – The training batch.

Returns:

The processed targets tensor.

Return type:

torch.Tensor