l^(hZddlmZddlmZddlmZddlmZddlm Z m Z Gdde Z y) ) annotations)Literal)nn) CrossEncoder) LambdaLossNoWeightingSchemeceZdZddddejdf dfd Zd fd Zed dZxZ S) RankNetLossNg?g|=binaryc Dt||t||||||y)a RankNet loss implementation for learning to rank. This loss function implements the RankNet algorithm, which learns a ranking function by optimizing pairwise document comparisons using a neural network. The implementation is optimized to handle padded documents efficiently by only processing valid documents during model inference. Args: model (CrossEncoder): CrossEncoder model to be trained sigma (float): Score difference weight used in sigmoid (default: 1.0) eps (float): Small constant for numerical stability (default: 1e-10) activation_fn (:class:`~torch.nn.Module`): Activation function applied to the logits before computing the loss. Defaults to :class:`~torch.nn.Identity`. mini_batch_size (int, optional): Number of samples to process in each forward pass. This has a significant impact on the memory consumption and speed of the training process. Three cases are possible: - If ``mini_batch_size`` is None, the ``mini_batch_size`` is set to the batch size. - If ``mini_batch_size`` is greater than 0, the batch is split into mini-batches of size ``mini_batch_size``. - If ``mini_batch_size`` is <= 0, the entire batch is processed at once. Defaults to None. References: - Learning to Rank using Gradient Descent: https://icml.cc/Conferences/2015/wp-content/uploads/2015/06/icml_ranking.pdf - `Cross Encoder > Training Examples > MS MARCO <../../../examples/cross_encoder/training/ms_marco/README.html>`_ Requirements: 1. Query with multiple documents (pairwise approach) 2. Documents must have relevance scores/labels. Both binary and continuous labels are supported. Inputs: +----------------------------------------+--------------------------------+-------------------------------+ | Texts | Labels | Number of Model Output Labels | +========================================+================================+===============================+ | (query, [doc1, doc2, ..., docN]) | [score1, score2, ..., scoreN] | 1 | +----------------------------------------+--------------------------------+-------------------------------+ Recommendations: - Use :class:`~sentence_transformers.util.mine_hard_negatives` with ``output_format="labeled-list"`` to convert question-answer pairs to the required input format with hard negatives. Relations: - :class:`~sentence_transformers.cross_encoder.losses.LambdaLoss` can be seen as an extension of this loss where each score pair is weighted. Alternatively, this loss can be seen as a special case of the :class:`~sentence_transformers.cross_encoder.losses.LambdaLoss` without a weighting scheme. - :class:`~sentence_transformers.cross_encoder.losses.LambdaLoss` with its default NDCGLoss2++ weighting scheme anecdotally performs better than the other losses with the same input format. Example: :: from sentence_transformers.cross_encoder import CrossEncoder, CrossEncoderTrainer, losses from datasets import Dataset model = CrossEncoder("microsoft/mpnet-base") train_dataset = Dataset.from_dict({ "query": ["What are pandas?", "What is the capital of France?"], "docs": [ ["Pandas are a kind of bear.", "Pandas are kind of like fish."], ["The capital of France is Paris.", "Paris is the capital of France.", "Paris is quite large."], ], "labels": [[1, 0], [1, 1, 0]], }) loss = losses.RankNetLoss(model) trainer = CrossEncoderTrainer( model=model, train_dataset=train_dataset, loss=loss, ) trainer.train() )modelweighting_schemeksigmaeps reduction_log activation_fnmini_batch_sizeN)super__init__r) selfr rrrrrr __class__s t/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/sentence_transformers/cross_encoder/losses/RankNetLoss.pyrzRankNetLoss.__init__ s3^ .0''+  c*t|}|d=|S)z Get configuration parameters for this loss function. Returns: Dictionary containing the configuration parameters r)rget_config_dict)rconfigrs rrzRankNetLoss.get_config_dictfs (* % & rcy)NaR @inproceedings{burges2005learning, title={Learning to Rank using Gradient Descent}, author={Burges, Chris and Shaked, Tal and Renshaw, Erin and Lazier, Ari and Deeds, Matt and Hamilton, Nicole and Hullender, Greg}, booktitle={Proceedings of the 22nd international conference on Machine learning}, pages={89--96}, year={2005} } )rs rcitationzRankNetLoss.citationqsr)r rr int | Nonerfloatrr"rzLiteral['natural', 'binary']rznn.Module | Nonerr!returnNone)r#z#dict[str, float | int | str | None])r#str) __name__ __module__ __qualname__rIdentityrrpropertyr __classcell__)rs@rr r s6>*5"++-&*X X  X  X  X 4 X (X $X  X t   rr N) __future__rtypingrtorchr#sentence_transformers.cross_encoderr*sentence_transformers.cross_encoder.lossesrrr rrrr1s""<Tp*pr