quaterion.loss.pairwise_loss module

class PairwiseLoss(distance_metric_name: Distance = Distance.COSINE)

Bases: SimilarityLoss

Base class for pairwise losses.

Parameters:

distance_metric_name – Name of the distance function, e.g., Distance.

forward(embeddings: Tensor, pairs: Tensor, labels: Tensor, subgroups: Tensor) → Tensor

Compute loss value.

Parameters:

• embeddings – shape: (batch_size, vector_length)

• pairs – shape: (2 * pairs_count,) - contains a list of known similarity pairs in batch

• labels – shape: (pairs_count,) - similarity of the pair

• subgroups – shape: (2 * pairs_count,) - subgroup ids of objects

Returns:

Tensor – zero-size tensor, loss value

training: bool