o ,hm`"@sdZddlmZmZmZddlZddlmZddlmZm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZddgZGd ddeZd d ed e d ed ed e d e_deedeedeedeedeededededededededededefddZdeedeedeedeedeededededededededededefd d!Ze ed" #  # # # #d&deedeedeedeedeeded$eedededededededededef d%dZdS)'z'Implementation for the RAdam algorithm.)castOptionalUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc _params_doc _to_scalar_use_grad_for_differentiable _view_as_real OptimizerParamsTRAdamradamcseZdZ     ddddddded eeefd eeefd ed ed ede edededeffddZ fddZ ddZ e dddZZS)rMbP?g?g+?:0yE>rFN)foreachmaximize capturabledifferentiableparamslrbetaseps weight_decaydecoupled_weight_decayrrrrc st|tr|dkrtdd|kstd|d|ks%td|d|dkr1dks;ntd|dd|dkrGdksQntd |dd|ks\td |t||||||| || d } t|| dS) NrzTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: ) r r!r"r#rrrr$r) isinstancernumel ValueErrordictsuper__init__) selfrr r!r"r#r$rrrrdefaults __class__L/var/www/html/scripts/venv/lib/python3.10/site-packages/torch/optim/radam.pyr, s0 zRAdam.__init__cst||jD]Y}|dd|dd|dd|dd|dd|dD]4}|j|g}t|dkrat|d sat |d }|drWtj |t |j d ntj |t d |d <q-q dS) NrrFrr$rrrstepdtypedevicer5) r+ __setstate__ param_groups setdefaultstategetlentorch is_tensorfloattensorr r6)r-r;grouppp_statestep_valr/r1r2r8Hs(          zRAdam.__setstate__c Csd}|dD]m}|jdurs|t|O}|||jjr!td||j|j|} t| dkr^|dr@tjdt |j dntj dt d | d <tj |tj d | d <tj |tj d | d <|| d || d || d q|S)NFrz'RAdam does not support sparse gradientsrrr1r4r%r7r3) memory_formatexp_avg exp_avg_sq)gradr> is_complexappend is_sparse RuntimeErrorr;r=zerosr r6rA zeros_likepreserve_format) r-rBparams_with_gradgradsexp_avgs exp_avg_sqs state_steps has_complexrCr;r1r1r2 _init_group\s2        zRAdam._init_groupc Cs|d}|dur!t |}Wdn1swY|jD]G}g}g}g}g}g}ttttf|d\} } |||||||} t|||||| | |d|d|d|d|d|d|d |d | d q$|S) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr!r r#r"rrrrr$) beta1beta2r r#r"rrrrr$rV) _cuda_graph_capture_health_checkr> enable_gradr9rtupler@rWr) r-closurelossrBrQrRrSrTrUrXrYrVr1r1r2r3sF   z RAdam.step)rrrrFN)__name__ __module__ __qualname__rrr@rr\boolrr,r8rWrr3 __classcell__r1r1r/r2rsH        (#aImplements RAdam algorithm. .. math:: \begin{aligned} &\rule{110mm}{0.4pt} \\ &\textbf{input} : \gamma \text{ (lr)}, \: \beta_1, \beta_2 \text{ (betas)}, \: \theta_0 \text{ (params)}, \:f(\theta) \text{ (objective)}, \: \lambda \text{ (weightdecay)}, \:\textit{maximize} \\ &\hspace{13mm} \epsilon \text{ (epsilon)}, \textit{decoupled\_weight\_decay} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0 \leftarrow 0 \text{ ( second moment)}, \\ &\hspace{18mm} \rho_{\infty} \leftarrow 2/(1-\beta_2) -1 \\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{6mm}\textbf{if} \: \textit{maximize}: \\ &\hspace{12mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{6mm}\textbf{else} \\ &\hspace{12mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{6mm} \theta_t \leftarrow \theta_{t-1} \\ &\hspace{6mm} \textbf{if} \: \lambda \neq 0 \\ &\hspace{12mm}\textbf{if} \: \textit{decoupled\_weight\_decay} \\ &\hspace{18mm} \theta_t \leftarrow \theta_{t} - \gamma \lambda \theta_{t} \\ &\hspace{12mm}\textbf{else} \\ &\hspace{18mm} g_t \leftarrow g_t + \lambda \theta_{t} \\ &\hspace{6mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{6mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{6mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\ &\hspace{6mm}\rho_t \leftarrow \rho_{\infty} - 2 t \beta^t_2 /\big(1-\beta_2^t \big) \\[0.1.ex] &\hspace{6mm}\textbf{if} \: \rho_t > 5 \\ &\hspace{12mm} l_t \leftarrow \frac{\sqrt{ (1-\beta^t_2) }}{ \sqrt{v_t} +\epsilon } \\ &\hspace{12mm} r_t \leftarrow \sqrt{\frac{(\rho_t-4)(\rho_t-2)\rho_{\infty}}{(\rho_{\infty}-4)(\rho_{\infty}-2) \rho_t}} \\ &\hspace{12mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t} r_t l_t \\ &\hspace{6mm}\textbf{else} \\ &\hspace{12mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t} \\ &\rule{110mm}{0.4pt} \\[-1.ex] &\bf{return} \: \theta_t \\[-1.ex] &\rule{110mm}{0.4pt} \\[-1.ex] \end{aligned} For further details regarding the algorithm we refer to `On the variance of the adaptive learning rate and beyond`_. This implementation provides an option to use either the original weight_decay implementation as in Adam (where the weight_decay is applied to the gradient) or the one from AdamW (where weight_decay is applied to the weight) through the decoupled_weight_decay option. When decoupled_weight_decay is set to False (default), it uses the original Adam style weight decay, otherwise, it uses the AdamW style which corresponds more closely to the `author's implementation`_ in the RAdam paper. Further information about decoupled weight decay can be found in `Decoupled Weight Decay Regularization`_. z Args: a lr (float, Tensor, optional): learning rate (default: 1e-3) betas (Tuple[float, float], optional): coefficients used for computing running averages of gradient and its square (default: (0.9, 0.999)) eps (float, optional): term added to the denominator to improve numerical stability (default: 1e-8) weight_decay (float, optional): weight decay (L2 penalty) (default: 0) decoupled_weight_decay (bool, optional): whether to decouple the weight decay as in AdamW to obtain RAdamW. If True, the algorithm does not accumulate weight decay in the momentum nor variance. (default: False) z a .. _On the variance of the adaptive learning rate and beyond: https://arxiv.org/abs/1908.03265 .. _author's implementation: https://github.com/LiyuanLucasLiu/RAdam .. _Decoupled Weight Decay Regularization: https://arxiv.org/abs/1711.05101 rrRrSrTrUrXrYr r#r"r$rrrrVc stjs t|}t|D]\}}| s||n|| }||}||||}tjsH| rHt}|jj |jj kr@|jj |vsHJd|dt |rat |}t |}t |}t |d7}| ri|nt |}|dkr| r}| d||n|j||d}||d| |j||d|dd||}d||||}dd|dd|||fdd }fd d }| rtd k||d }|j|||ddq d kr|j||||ddq |j||ddq dS)NIIf capturable=True, params and state_steps must be on supported devices: .rralpha)valuecs,dddddS)Nrj?r1r1)rho_infrho_tr1r2 _compute_rectDsz+_single_tensor_radam.._compute_rectcs.}r |}n|}d|S)Nrl)sqrtaddadd_)exp_avg_sq_sqrt)bias_correction2rr"rHr1r2_compute_adaptive_lrLs    z2_single_tensor_radam.._compute_adaptive_lr@r&g)r>jit is_scriptingr enumeratecompiler is_compilingr r6typerJ view_as_realrmul_rqlerp_addcmul_whererr)rrRrSrTrUrXrYr r#r"r$rrrrViparamrIrGstep_tcapturable_supported_devicesr3bias_correction1bias_corrected_exp_avgroruupdater1)rtrr"rHrmrnr2_single_tensor_radamsf          rc %s8t|dkrdS| rJdtjs0| r0tddtfddt||Ds0Jddtt |||||g}| D]\\}}}}}}t t t |}t t t |}t t t |}t t t |}t t t |}tjs|djrtj|tjd d d d d nt|d |rt||||| rt|}dd d | rt|}t|t|d t|}t||t|dt||t|t||}n fdd|D}|dkr| rt|d |n| rtj|||d ntj|||d }t||d t|t|||d ~| rt|d}t|d}t||~t|ddt|} t|| ~ t|ddt||D}!~~dd|!D}"t|"t|}t|t|d t|"|t|"t|}t|t|d t|t|t||!~!t|t||~n3fdd|D}!dd|!D}#fdd|D}fddt|#|D}"fddt||!|D}t|}$t|$| t|$|t|$t|$|"t|||$qBdS)Nrz#_foreach ops don't support autogradF) supports_xlac3s0|]\}}|jj|jjko|jjvVqdSr_)r6r|).0rCr3)rr1r2 s   z&_multi_tensor_radam..rerfr&cpu)r6rgrrjcs8g|]}dt|t|dt|qS)rjrrrr3)rYrmr1r2 s z'_multi_tensor_radam..rkcSs"g|] \}}t|dk|dqS)rvr%r>r)rnrnr1r1r2rscSsg|] }t|dkddqS)rr%r&rrrectr1r1r2rscsDg|]}|dkr|d|ddd|dndqS)rkrjrlrr1)rrn)rmr1r2r s cSsg|] }|dkr dndqS)rr&r1rr1r1r2rscsg|] }dt|qS)rrr)rXr1r2rscs g|] \}}||dqS)r1)rrbc)r r1r2rscs6g|]\}}}dt|d||dqS)rrlrr)rr3rr)rYr r1r2r s")r=r>rzr{r allziprr"_group_tensors_by_device_and_dtypevaluesrlistris_cpu _foreach_add_rAr _foreach_neg _foreach_pow _foreach_neg_ _foreach_mul_ _foreach_div_ _foreach_add_foreach_lerp__foreach_addcmul_ _foreach_sub _foreach_mul_foreach_sqrt_ _foreach_sqrt_foreach_reciprocal_)%rrRrSrTrUrXrYr r#r"r$rrrrVgrouped_tensorsgrouped_params_grouped_grads_grouped_exp_avgs_grouped_exp_avg_sqs_grouped_state_steps__grouped_params grouped_gradsgrouped_exp_avgsgrouped_exp_avg_sqsgrouped_state_stepsrrt rho_t_listnumsub2denomrunrect_step_size unrectifiedbufferr1)rXrYrr rmr2_multi_tensor_radamhs                                                     r)single_tensor_fnFrc Cstdd|Ds td|durt||dd\}}|r%tjr%td|r/tjs/t}nt}||||||| | | ||| |||| ddS) zpFunctional API that performs RAdam algorithm computation. See :class:`~torch.optim.RAdam` for details. css|] }t|tjVqdSr_)r'r>r)rtr1r1r2rIszradam..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF) use_fusedz6torch.jit.script not supported with foreach optimizers) rXrYr r#r"rr$rrrV)rrMrr>rwrxrr)rrRrSrTrUr$rrrrVrrXrYr r#r"rfuncr1r1r2r/s<  )FNFFFF)__doc__typingrrrr>r optimizerrrr r r r r rrrrrrrr__all__rrr@rcrrrr1r1r1r2s D3  P      h      H