quaterion.loss.contrastive_loss module

class ContrastiveLoss(distance_metric_name: Distance = Distance.COSINE, margin: float = 0.5, size_average: bool = True)

Bases: PairwiseLoss

Contrastive loss.

Expects as input two texts and a label of either 0 or 1. If the label == 1, then the distance between the two embeddings is reduced. If the label == 0, then the distance between the embeddings is increased.

Further information:

http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf

Parameters:

• distance_metric_name – Name of the function, e.g., Distance. Optional, defaults to COSINE.

• margin – Negative samples (label == 0) should have a distance of at least the margin value.

• size_average – Average by the size of the mini-batch.

forward(embeddings: Tensor, pairs: LongTensor, labels: Tensor, subgroups: Tensor, **kwargs) → Tensor

Compute loss value.

Parameters:

• embeddings – Batch of embeddings, first half of embeddings are embeddings of first objects in pairs, second half are embeddings of second objects in pairs.

• pairs – Indices of corresponding objects in pairs.

• labels – Scores of positive and negative objects.

• subgroups – subgroups to distinguish objects which can and cannot be used as negative examples

• **kwargs – additional key-word arguments for generalization of loss call

Returns:

Tensor – averaged or summed loss value

get_config_dict() → Dict[str, Any]

Config used in saving and loading purposes.

Config object has to be JSON-serializable.

Returns:

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

training: bool