quaterion.loss.triplet_loss module

class TripletLoss(margin: float | None = 0.5, distance_metric_name: Distance | None = Distance.COSINE, mining: str | None = 'hard', soft: bool | None = False)

Bases: GroupLoss

Implements Triplet Loss as defined in https://arxiv.org/abs/1503.03832

It supports batch-all, batch-hard and batch-semihard strategies for online triplet mining.

Parameters:

• margin – Margin value to push negative examples apart.

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

• mining – Triplet mining strategy. One of “all”, “hard”, “semi_hard”.

• soft – If True, use soft margin variant of Hard Triplet Loss. Ignored in all other cases.

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

Calculates Triplet Loss with specified embeddings and labels.

Parameters:

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

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

Returns:

torch.Tensor – Scalar loss value.

get_config_dict()

Config used in saving and loading purposes.

Config object has to be JSON-serializable.

Returns:

Dict[str, Any] – JSON-serializable dict of params

xbm_loss(embeddings: Tensor, groups: LongTensor, memory_embeddings: Tensor, memory_groups: LongTensor) → Tensor

Implement XBM loss computation for this loss.

Parameters:

• embeddings – shape: (batch_size, vector_length) - Output embeddings from the encoder.

• groups – shape: (batch_size,) - Group ids associated with embeddings.

• memory_embeddings – shape: (memory_buffer_size, vector_length) - Embeddings stored in a ring buffer

• memory_groups – shape: (memory_buffer_size,) - Groups ids associated with memory_embeddings

Returns:

Tensor – zero-size tensor, XBM loss value.

training: bool