^j$dZddlZddlmZmZddlZddlmZddlm Z dejdejd ejd ejd e d e d e de dejde de de de dedededee e fdef$dZGddeZy)a PyTorch MARS Optimizer Code simplified from https://github.com/AGI-Arena/MARS Paper: MARS: Unleashing the Power of Variance Reduction for Training Large Models - https://arxiv.org/abs/2411.10438 @article{yuan2024mars, title={MARS: Unleashing the Power of Variance Reduction for Training Large Models}, author={Yuan, Huizhuo and Liu, Yifeng and Wu, Shuang and Zhou, Xun and Gu, Quanquan}, journal={arXiv preprint arXiv:2411.10438}, year={2024} } N)OptionalTuple) Optimizer)ParamsTpgradexp_avg exp_avg_sqlr weight_decaybeta1beta2 last_gradepsstepgamma mars_type is_grad_2d optimize_1d lr_1d_factorbetas_1dcautioncl|s| rd|z }| dk(r|}nH||z j| ||z zj|}tj|}|dkDr||z }|j|j|||rU||zdkDj |j }|j |jjd||z}| dk(r|j|j||d|z d|| zz }d|| zz }|jtj|z j| }||z||z j |z}n| d k(r||z|jz}nJ|j|| ||fS|\}}|j|j|d|z |j|j||d|z d|| zz }d|| zz }|jtj|z j| }|rU||zdkDj |j }|j |jjd||z}||z||z j |z}|j|||z ||fS) N?r)alphargMbP?)minadamw)valuelion) mul_add_torchnormtodtypediv_meanclamp_addcmul_sqrtmathsign)rr r r r r rrrrrrrrrrrrone_minus_beta1c_tc_t_normmaskbias_correction1bias_correction2denomupdatebeta1_1dbeta2_1ds Z/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/timm/optim/mars.py_mars_single_tensor_stepr9s*ju* 19C)#))%5?3J*KLQQRVWCzz#H"}Hn U  O < dNQ&**4::6D IIdiik((T(2 3nG   OOE " + +CBJ + G"Ud]2 "Ud]2 __&3C)DDJJ3OE%3C)C(I(I%(PPF & % 6F 5 vbS! J &( X##DX #>!**4R(]*KT!11T!11"TYY/?%@@FFsK dNQ&**4::6D IIdiik((T(2 3nG\!W/?%?$E$Ee$LL vrL012 J ceZdZdZ ddededeeefdedededed ed ed e eeefd effd Z fdZ e jddZxZS)Marsz MARS Optimizer Paper: MARS: Unleashing the Power of Variance Reduction for Training Large Models https://arxiv.org/abs/2411.10438 paramsr betasrr rrrrrrc d|kstdj|d|kstdj|d|dcxkrdksntdj|dd|dcxkrdksntdj|d|d vsJd t|||||||| | xs|| } tt||| y) NzInvalid learning rate: {}zInvalid epsilon value: {}rrz%Invalid beta parameter at index 0: {}rz%Invalid beta parameter at index 1: {})rr zMARS type not supported) r r>rr rrrrrr) ValueErrorformatdictsuperr<__init__)selfr=r r>rr rrrrrrdefaults __class__s r8rEz Mars.__init__bsby8??CD Dcz8??DE EeAh$$DKKERSHUV VeAh$$DKKERSHUV V--H/HH-%#%&   dD"684r:crtt| ||jD]}|j ddy)NrF)rDr< __setstate__ param_groups setdefault)rFstategrouprHs r8rJzMars.__setstate__s7 dD&u-&& /E   Y . /r:cd}|$tj5|}ddd|jD]"}|dD]}|j|j}|jr t d|j |}t|dkrMd|d<tj||d<tj||d<tj||d <|dxxdz cc<|d}|d}|d } |d} |d } |d } |d \} }|jd k\}t|||| | | | || |d||d|d||d|d|d|d||d<%|S#1swY>xYw)zPerforms a single optimization step. Arguments: closure (callable, optional): A closure that reevaluates the model and returns the loss. Nr=zJAdam does not support sparse gradients, please consider SparseAdam insteadrrrr rr r r r>rrrrrrr)rrrrrr) r# enable_gradrKr is_sparse RuntimeErrorrMlen zeros_likendimr9)rFclosurelossrNrr rMrr r rr wdrrrs r8rz Mars.steps  ""$ !y !&&3 *E8_2 *66>vv>>&'stt 1 u:?$%E&M','7'7':E)$).)9)9!)*$W~ u!YY!^ )%L'N#K0) %m 4!&~!6":.!),%*&*k"e2 *3 *j q ! !s EE') g~jth?)g?gGz?g:0yE>r@g?rFrNF)N)__name__ __module__ __qualname____doc__rfloatrstrboolrrErJr#no_gradr __classcell__)rHs@r8r<r<[s)4"$ $ %"%6:!$5$5$5& $5  $5  $5$5$5$5 $5uUE\23$5$5L/ U]]_AAr:r<)r]r,typingrrr#torch.optim.optimizerr_typesrTensorr^intr_r`r9r<r:r8ris  " +? <<?ll??LL ?  ?  ???<<??????? !?"u %#?$%?Dt9tr: