h jSSKJr SSKJr SSKrSSKJrJr SSKJr "SS\R5r g)) annotations)IterableN)Tensornn) SparseEncodercX^\rSrSrSSU4SjjjrS SjrS Sjr\S Sj5rSr U=r $) FlopsLoss c:>[TU]5 XlX lg)aO FlopsLoss implements a regularization technique to promote sparsity in sparse encoder models. It calculates the squared L2 norm of the mean embedding vector, which helps reduce the number of floating-point operations (FLOPs) required during inference by encouraging more zero values in the embeddings. It can use a threshold to ignore embeddings with too few non-zero (active) elements. This loss is used as a regularization component within other losses like :class:`SpladeLoss` rather than being used as a standalone loss function. Args: model: SparseEncoder model to be regularized threshold: Optional threshold for the number of non-zero (active) elements in the embeddings. If specified, only embeddings with more than this number of non-zero (active) elements will be considered. This can help to ignore embeddings that are too sparse and may not contribute meaningfully to the loss. References: - For further details, see: https://arxiv.org/pdf/2004.05665 for the general FLOPS loss and https://arxiv.org/pdf/2504.14839 for FLOPS with thresholds, a.k.a. FLOPS with l0 masking. Relations: - Used as a component within :class:`SpladeLoss` to regularize both query and document embeddings Example: - This loss is typically used within the :class:`SpladeLoss` class, which combines it with other loss components. N)super__init__model threshold)selfrr __class__s o/home/james-whalen/.local/lib/python3.13/site-packages/sentence_transformers/sparse_encoder/losses/FlopsLoss.pyr FlopsLoss.__init__ s4  "c[S5e)NzaFlopsLoss is not intended to be used directly. Use it as a regulizer within the SpladeLoss class.)NotImplementedError)rsentence_featureslabelss rforwardFlopsLoss.forward*s! o  rcURbAUS:gRSS9nX R:R5nXRS5-n[R"[R "USS9S-5$)Nr)dim)rsumfloat unsqueezetorchmean)r embeddingsl0_normmasks rcompute_loss_from_embeddings&FlopsLoss.compute_loss_from_embeddings/sh >> %!Q+++2Gnn,335D#nnQ&77JyyJA6!;<r;s$"$ L5 5r