quaterion.loss.circle_loss module

class CircleLoss(margin: float | None = 0.25, scale_factor: float | None = 256, distance_metric_name: Distance | None = Distance.COSINE)

Bases: GroupLoss

Implements Circle Loss as defined in https://arxiv.org/abs/2002.10857.

Parameters:

• margin – Margin value to push negative examples.

• scale_factor – scale factor γ determines the largest scale of each similarity score.

Note

Refer to sections 4.1 and 4.5 in the paper for default values and evaluation of margin and scaling_factor hyperparameters.

forward(embeddings: Tensor, groups: LongTensor) → Tensor

Compute loss value.

Parameters:

• embeddings – shape: (batch_size, vector_length) - Batch of embeddings.

• groups – shape: (batch_size,) - Batch of labels associated with embeddings

Returns:

Tensor – Scalar loss value.

training: bool