l^(h8ddlmZddlmZmZddlmZddlmZddl m Z ddl Z ddl Z ddl m Z mZddlmZmZdd lmZmZdd lmZGd d Z ddZGdd ej2Zy)) annotations)IterableIterator) nullcontext)partial)AnyN)Tensornn)get_device_statesset_device_states)SentenceTransformerutil)StaticEmbeddingc(eZdZdZddZddZddZy) RandContexta Random-state context manager class. Reference: https://github.com/luyug/GradCache. This class will back up the pytorch's random state during initialization. Then when the context is activated, the class will set up the random state with the backed-up one. c`tj|_t|\|_|_yN)torch get_rng_state fwd_cpu_stater fwd_gpu_devicesfwd_gpu_states)selftensorss }/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/sentence_transformers/losses/CachedMultipleNegativesRankingLoss.py__init__zRandContext.__init__s("0024Ew4O1d1ctjj|jd|_|jj tj |jt|j|jy)NT)devicesenabled) rrandomfork_rngr_fork __enter__ set_rng_staterr rrs rr$zRandContext.__enter__s^\\**43G3GQU*V   D../$..0C0CDrcL|jj|||d|_yr)r#__exit__)rexc_typeexc_valexc_tbs rr(zRandContext.__exit__#s Hgv6 rN)returnNone)__name__ __module__ __qualname____doc__rr$r(rrrrsPE rr"CachedMultipleNegativesRankingLossc |jJ|jJtj5t ||j|jD]u\}}}t |j |dd||D]M\\}}}tj |j|j|z} | jOw dddy#1swYyxYw)zOA backward hook to backpropagate the cached gradients mini-batch by mini-batch.NTF)sentence_feature with_gradcopy_random_state random_states) cacher8r enable_gradzipembed_minibatch_iterdotflattenbackward) grad_outputsentence_featuresloss_objr5gradr8reps_mb_grad_mb surrogates r_backward_hookrH(s >> %% %  ! ! -- -     %589JHNN\d\r\r5s % 1 dM),--%5"&+"/ . * %% !g"IIgoo&79JKkY ""$ % % % % %s BCCceZdZdejddf d fd Z d ddZ d ddZddZdddZ dd Z dd Z e dd Z xZS)r3g4@ Fct|t|dtr t d||_||_||_tj|_ ||_ d|_ d|_ ||_y)a Boosted version of MultipleNegativesRankingLoss (https://arxiv.org/pdf/1705.00652.pdf) by GradCache (https://arxiv.org/pdf/2101.06983.pdf). Constrastive learning (here our MNRL loss) with in-batch negatives is usually hard to work with large batch sizes due to (GPU) memory limitation. Even with batch-scaling methods like gradient-scaling, it cannot work either. This is because the in-batch negatives make the data points within the same batch non-independent and thus the batch cannot be broke down into mini-batches. GradCache is a smart way to solve this problem. It achieves the goal by dividing the computation into two stages of embedding and loss calculation, which both can be scaled by mini-batches. As a result, memory of constant size (e.g. that works with batch size = 32) can now process much larger batches (e.g. 65536). In detail: (1) It first does a quick embedding step without gradients/computation graphs to get all the embeddings; (2) Calculate the loss, backward up to the embeddings and cache the gradients wrt. to the embeddings; (3) A 2nd embedding step with gradients/computation graphs and connect the cached gradients into the backward chain. Notes: All steps are done with mini-batches. In the original implementation of GradCache, (2) is not done in mini-batches and requires a lot memory when the batch size is large. One drawback is about the speed. Gradient caching will sacrifice around 20% computation time according to the paper. Args: model: SentenceTransformer model scale: Output of similarity function is multiplied by scale value similarity_fct: similarity function between sentence embeddings. By default, cos_sim. Can also be set to dot product (and then set scale to 1) mini_batch_size: Mini-batch size for the forward pass, this denotes how much memory is actually used during training and evaluation. The larger the mini-batch size, the more memory efficient the training is, but the slower the training will be. It's recommended to set it as high as your GPU memory allows. The default value is 32. show_progress_bar: If True, a progress bar for the mini-batches is shown during training. The default is False. References: - Efficient Natural Language Response Suggestion for Smart Reply, Section 4.4: https://arxiv.org/pdf/1705.00652.pdf - Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup: https://arxiv.org/pdf/2101.06983.pdf Requirements: 1. (anchor, positive) pairs or (anchor, positive, negative pairs) 2. Should be used with large `per_device_train_batch_size` and low `mini_batch_size` for superior performance, but slower training time than :class:`MultipleNegativesRankingLoss`. Inputs: +-------------------------------------------------+--------+ | Texts | Labels | +=================================================+========+ | (anchor, positive) pairs | none | +-------------------------------------------------+--------+ | (anchor, positive, negative) triplets | none | +-------------------------------------------------+--------+ | (anchor, positive, negative_1, ..., negative_n) | none | +-------------------------------------------------+--------+ Recommendations: - Use ``BatchSamplers.NO_DUPLICATES`` (:class:`docs `) to ensure that no in-batch negatives are duplicates of the anchor or positive samples. Relations: - Equivalent to :class:`MultipleNegativesRankingLoss`, but with caching that allows for much higher batch sizes (and thus better performance) without extra memory usage. This loss also trains roughly 2x to 2.4x slower than :class:`MultipleNegativesRankingLoss`. Example: :: from sentence_transformers import SentenceTransformer, SentenceTransformerTrainer, losses from datasets import Dataset model = SentenceTransformer("microsoft/mpnet-base") train_dataset = Dataset.from_dict({ "anchor": ["It's nice weather outside today.", "He drove to work."], "positive": ["It's so sunny.", "He took the car to the office."], }) loss = losses.CachedMultipleNegativesRankingLoss(model, mini_batch_size=64) trainer = SentenceTransformerTrainer( model=model, train_dataset=train_dataset, loss=loss, ) trainer.train() rzCachedMultipleNegativesRankingLoss is not compatible with a SentenceTransformer model based on a StaticEmbedding. Consider using MultipleNegativesRankingLoss instead.N)superr isinstancer ValueErrormodelscalesimilarity_fctr CrossEntropyLosscross_entropy_lossmini_batch_sizer9r8show_progress_bar)rrOrPrQrTrU __class__s rrz+CachedMultipleNegativesRankingLoss.__init__@szl  eAh 0G    ,"$"5"5"7.04 =A!2rcv|rtntj}| tn|}|jD cic] \} } | | || } } } |5|5|rt | j nd}|j | d} dddddd |fScc} } w#1swYxYw#1swY |fSxYw)zYDo forward pass on a minibatch of the input features and return corresponding embeddings.Nsentence_embedding)rrno_graditemsrvaluesrO) rr5beginendr6r7 random_state grad_contextrandom_state_contextkvsentence_feature_minibatchrepss rembed_minibatchz2CachedMultipleNegativesRankingLoss.embed_minibatchs'0{U]] 0<0D{},BRBXBXBZ%[$!Qa5o%["%[ ! T TTe{,F,M,M,OPko zz"<=>RS T T\!! &\ T T T\!!s)BB,/B B, B) %B,,B8c # K|d}|j\}}ttjd||jd|j D];\}} | |jz} |j || | |||dn||\} } | | f=yw)z`Do forward pass on all the minibatches of the input features and yield corresponding embeddings. input_idsrzEmbed mini-batchesdescdisableN)r5r\r]r6r7r^)shape enumeratetqdmtrangerTrUre) rr5r6r7r8rgbszrEiberdr^s rr<z7CachedMultipleNegativesRankingLoss.embed_minibatch_iters-[9 Q KK$$) 222    %DAqD(((A!%!5!5!1#"3%2%:T a@P "6" D, $ $% %sB Bc|j|d}|jj}|Dcgc]}|Dcgc]}|jc}c}}|_|Scc}wcc}}w)zMCalculate the cross-entropy loss and cache the gradients wrt. the embeddings.T) with_backward)calculate_lossdetachrequires_grad_rCr9)rrdlossrsrs r"calculate_loss_and_cache_gradientszECachedMultipleNegativesRankingLoss.calculate_loss_and_cache_gradientssY""4t"<{{}++-59:rr*!qvv*:  +:s A(A# A(#A(ctj|d}tj|ddDcgc]}tj|c}}t|}tjt |tj |j }g}tjd||jd|j D]} | |jz} |j|| | ||jz} |j| || | t| z|z } |r | j| j} |j!| t#|} | Scc}w)zACalculate the cross-entropy loss. No need to cache the gradients.rN)dtypedevicezPreparing cachesrh)rcatlentensorrangelongrrmrnrTrUrQrPrSr?rvappendsum)rrdrt embeddings_arz embeddings_b batch_sizelabelslossesrqrrscores loss_mbatchrxs rruz1CachedMultipleNegativesRankingLoss.calculate_losssGyya) yyQR!A1%))A,!AB &  * UZZ 8K8K &(   #...  'AD(((A!00a1BLQTXT^T^^F(,(?(?qQR (TWZ[aWb(beo(oK$$&)002 MM+ & '6{ /"BsE!cg}g|_|D]}g}g}|j|ddD]C\}}|j|jj |j|E|j||jj|t j r4|j|} | jtt||| S|j|} | S)NFT)r5r6r7)rArB) r8r<rrvrwris_grad_enabledr{ register_hookrrHru) rrArrdr5reps_mbsrandom_state_mbsrDr^rxs rforwardz*CachedMultipleNegativesRankingLoss.forwards 1 8 H! )-)B)B!1"&*C* 6%  0 ? ? AB '' 5  6 KK !    % %&6 7 8  "::4@D   w~IZeij k  &&t,D rc^|j|jj|jdS)N)rPrQrT)rPrQr.rTr&s rget_config_dictz2CachedMultipleNegativesRankingLoss.get_config_dicts+ZZ"11::#33  rcy)Na @misc{gao2021scaling, title={Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup}, author={Luyu Gao and Yunyi Zhang and Jiawei Han and Jamie Callan}, year={2021}, eprint={2101.06983}, archivePrefix={arXiv}, primaryClass={cs.LG} } r2r&s rcitationz+CachedMultipleNegativesRankingLoss.citation s r) rOr rPfloatrQz"callable[[Tensor, Tensor], Tensor]rTintrUboolr,r-r)r5dict[str, Tensor]r\rr]rr6rr7rr^zRandContext | Noner,z!tuple[Tensor, RandContext | None]) r5rr6rr7rr8zlist[RandContext] | Noner,z+Iterator[tuple[Tensor, RandContext | None]])rdlist[list[Tensor]]r,r )F)rdrrtrr,r )rAIterable[dict[str, Tensor]]rr r,r )r,zdict[str, Any])r,str)r.r/r0rcos_simrrer<r{rurrpropertyr __classcell__)rVs@rr3r3?s =A\\!"' d3"d3d3; d3  d3  d3 d3Z,0"+"" "  "  ")" +"037 %+%% % 0 % 5 %<8:   r)r@r rArrBr3r,r-) __future__rcollections.abcrr contextlibr functoolsrtypingrrrmr r torch.utils.checkpointr r sentence_transformersr rsentence_transformers.modelsrrrHModuler3r2rrrsk"." G;8.%%2%1% %.llr