Contrastive Loss API¶

Bases: Module

Hyperbolic InfoNCE loss operating directly on Lorentz-model embeddings.

The encoder now returns hyperbolic embeddings directly, so this loss function works with them without additional projection.

`forward(anchor_emb, positive_emb, negative_embs, batch_size, k_negatives, false_negative_mask=None, adaptive_margins=None)` ¶

Compute Hyperbolic InfoNCE loss.

Parameters:

Name	Type	Description	Default
`anchor_emb`	`Tensor`	Anchor hyperbolic embeddings (batch_size, embedding_dim+1)	required
`positive_emb`	`Tensor`	Positive hyperbolic embeddings (batch_size, embedding_dim+1)	required
`negative_embs`	`Tensor`	Negative hyperbolic embeddings (batch_size * k_negatives, embedding_dim+1)	required
`batch_size`	`int`	Batch size	required
`k_negatives`	`int`	Number of negatives per anchor	required
`false_negative_mask`	`Optional[Tensor]`	Optional mask for false negatives (batch_size, k_negatives)	`None`
`adaptive_margins`	`Optional[Tensor]`	Optional per-anchor adaptive margins (batch_size,)	`None`

Returns:

Type	Description
`Tensor`	Loss scalar