rh`%jddlmZddlZddlmZmZddlmZddlmZGddejZ y)) annotationsN)Tensornn) CrossEncoder)fullnamecjeZdZejfdfd ZddZddZdZe d dZ xZ S) MarginMSELossc t|||_||_t j di||_t|jts8t|jjdt|jd|jjdk7r9t|jjd|jjdy)a Computes the MSE loss between ``|sim(Query, Pos) - sim(Query, Neg)|`` and ``|gold_sim(Query, Pos) - gold_sim(Query, Neg)|``. This loss is often used to distill a cross-encoder model from a teacher cross-encoder model or gold labels. In contrast to :class:`~sentence_transformers.cross_encoder.losses.MultipleNegativesRankingLoss`, the two passages do not have to be strictly positive and negative, both can be relevant or not relevant for a given query. This can be an advantage of MarginMSELoss over MultipleNegativesRankingLoss. .. note:: Be mindful of the magnitude of both the labels and what the model produces. If the teacher model produces logits with Sigmoid to bound them to [0, 1], then you may wish to use a Sigmoid activation function in the loss. Args: model (:class:`~sentence_transformers.cross_encoder.CrossEncoder`): A CrossEncoder model to be trained. activation_fn (:class:`~torch.nn.Module`): Activation function applied to the logits before computing the loss. **kwargs: Additional keyword arguments passed to the underlying :class:`torch.nn.MSELoss`. References: - Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation: https://arxiv.org/abs/2010.02666 - `Cross Encoder > Training Examples > Distillation <../../../examples/cross_encoder/training/distillation/README.html>`_ Requirements: 1. Your model must be initialized with `num_labels = 1` (a.k.a. the default) to predict one class. 2. Usually uses a finetuned CrossEncoder teacher M in a knowledge distillation setup. Inputs: +------------------------------------------------+--------------------------------------------------------------------------------------------+-------------------------------+ | Texts | Labels | Number of Model Output Labels | +================================================+============================================================================================+===============================+ | (query, passage_one, passage_two) triplets | gold_sim(query, passage_one) - gold_sim(query, passage_two) | 1 | +------------------------------------------------+--------------------------------------------------------------------------------------------+-------------------------------+ | (query, passage_one, passage_two) triplets | [gold_sim(query, passage_one), gold_sim(query, passage_two)] | 1 | +------------------------------------------------+--------------------------------------------------------------------------------------------+-------------------------------+ | (query, positive, negative_1, ..., negative_n) | [gold_sim(query, positive) - gold_sim(query, negative_i) for i in 1..n] | 1 | +------------------------------------------------+--------------------------------------------------------------------------------------------+-------------------------------+ | (query, positive, negative_1, ..., negative_n) | [gold_sim(query, positive), gold_sim(query, negative_1), ..., gold_sim(query, negative_n)] | 1 | +------------------------------------------------+--------------------------------------------------------------------------------------------+-------------------------------+ Relations: - :class:`MSELoss` is similar to this loss, but without a margin through the negative pair. Example: :: from sentence_transformers.cross_encoder import CrossEncoder, CrossEncoderTrainer, losses from datasets import Dataset student_model = CrossEncoder("microsoft/mpnet-base") teacher_model = CrossEncoder("cross-encoder/ms-marco-MiniLM-L12-v2") train_dataset = Dataset.from_dict({ "query": ["What are pandas?", "What is the capital of France?"], "positive": ["Pandas are a kind of bear.", "The capital of France is Paris."], "negative": ["Pandas are a kind of fish.", "The capital of France is Berlin."], }) def compute_labels(batch): positive_scores = teacher_model.predict(list(zip(batch["query"], batch["positive"]))) negative_scores = teacher_model.predict(list(zip(batch["query"], batch["negative"]))) return { "label": positive_scores - negative_scores } train_dataset = train_dataset.map(compute_labels, batched=True) loss = losses.MarginMSELoss(student_model) trainer = CrossEncoderTrainer( model=student_model, train_dataset=train_dataset, loss=loss, ) trainer.train() z? expects a model of type CrossEncoder, but got a model of type .z; expects a model with 1 output label, but got a model with z output labels.N)super__init__model activation_fnrMSELossloss_fct isinstancer ValueError __class____name__type num_labels)selfrrkwargsrs /var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/sentence_transformers/cross_encoder/losses/MarginMSELoss.pyrzMarginMSELoss.__init__ sT  * ,V, $**l3>>**+,++/ +;*>**+,((, (=(='>oO  &c $|d}|d}|dd}t|}t|trtj|}|j |t|dzfk(r$|dddfj d|ddddfz }|j |fk(r|j d}|j |t|fk7r'td|j d|t|fdtt||}|j|}g} |D]7} tt|| } | j|j| 9|j dtj| dz } |j| |j} | S)Nrr z Labels shape z does not match expected shape z. Ensure that your dataset labels/scores are 1) lists of differences between positive scores and negatives scores (length `num_negatives`), or 2) lists of positive and negative scores (length `num_negatives + 1`).)dim) lenrlisttorchstackshape unsqueezerziplogits_from_pairsappendrfloat)rinputslabelsanchors positives negatives batch_sizepositive_pairspositive_logitsnegative_logits_listnegativenegative_pairs margin_logitslosss rforwardzMarginMSELoss.forwardfs)1I 12J \  fd #[[(F <F <(>~(N O P(11!4u{{CW]^7__ }}]FLLN; rc|jj|ddd}|j|jj|jdi|dj d}|j |S)z Computes the logits for a list of pairs using the model. Args: pairs (list[tuple[str, str]]): A list of pairs of strings (query, passage). Returns: Tensor: The logits for the pairs. Tpt)padding truncationreturn_tensorsrr )r tokenizertodeviceviewr)rpairstokenslogitss rr(zMarginMSELoss.logits_from_pairssu%%  &   $**##$%f%a(--b1!!&))rc0dt|jiS)Nr)rrrs rget_config_dictzMarginMSELoss.get_config_dicts Xd&8&89  rcy)NuY @misc{hofstätter2021improving, title={Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation}, author={Sebastian Hofstätter and Sophia Althammer and Michael Schröder and Mete Sertkan and Allan Hanbury}, year={2021}, eprint={2010.02666}, archivePrefix={arXiv}, primaryClass={cs.IR} } r rGs rcitationzMarginMSELoss.citations r)rrrz nn.ModulereturnNone)r+zlist[list[str]]r,zTensor | list[Tensor]rKr)rCzlist[tuple[str, str]]rKr)rKstr) r __module__ __qualname__rIdentityrr8r(rHpropertyrJ __classcell__)rs@rr r s9GRr{{}Yv%N*(   rr ) __future__rr#rr0sentence_transformers.cross_encoder.CrossEncoderrsentence_transformers.utilrModuler r rrrWs&" I/gBIIgr