a hm`@sdZddlmZmZmZddlZddlmZddlmZm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZddgZGd ddeZd d ed e d ed ed e d e_eeeeeeeeeeeeeeeeeeeedddZeeeeeeeeeeeeeeeeeeeedddZe eddeeeeeeeeeeeeeeeeeeeeeedddZdS)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 OptimizerParamsTRAdamradamc sxeZdZddddddeeeefeeefeeee eeeed fd d Z fd d Z d dZ e dddZZS)rMbP?g?g+?:0yE>rFN)foreachmaximize capturabledifferentiable) paramslrbetaseps weight_decaydecoupled_weight_decayrrrrc st|tr|dkrtdd|ks4td|d|ksJtd|d|dkrbdksvntd|dd|dkrdksntd |dd|kstd |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$r) isinstancerZnumel ValueErrordictsuper__init__) selfrr r!r"r#r$rrrrdefaults __class__?/var/www/auris/lib/python3.9/site-packages/torch/optim/radam.pyr+ s0 zRAdam.__init__cst||jD]}|dd|dd|dd|dd|dd|dD]h}|j|g}t|dkrZt|d sZt |d }|drtj |t |j d ntj |t d |d <qZqdS) NrrFrr$rrrstepdtypedevicer4) r* __setstate__ param_groups setdefaultstategetlentorchZ is_tensorfloattensorr r5)r,r:grouppZp_stateZstep_valr.r0r1r7Hs"          zRAdam.__setstate__c Csd}|dD]}|jdur |t|O}|||jjrBtd||j|j|} t| dkr|drtjdt |j dntj dt d | d <tj |tj d | d <tj |tj d | d <|| d || d || d q |S)NFrz'RAdam does not support sparse gradientsrrr0r3r%r6r2)Z memory_formatexp_avg exp_avg_sq)gradr= is_complexappendZ is_sparse RuntimeErrorr:r<zerosr r5r?Z zeros_likeZpreserve_format) r,r@params_with_gradgradsexp_avgs exp_avg_sqs state_steps has_complexrAr:r0r0r1 _init_group\s0        zRAdam._init_groupc Cs|d}|durBt|}Wdn1s80Y|jD]}g}g}g}g}g}ttttf|d\} } |||||||} t|||||| | |d|d|d|d|d|d|d |d | d qH|S) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr!r r#r"rrrrr$) beta1beta2r r#r"rrrrr$rN) Z _cuda_graph_capture_health_checkr=Z enable_gradr8rtupler>rOr) r,closureZlossr@rIrJrKrLrMrPrQrNr0r0r1r2sD $  z RAdam.step)rrrrF)N)__name__ __module__ __qualname__rrr>rrRboolrr+r7rOrr2 __classcell__r0r0r.r1rs4   ( #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 )rrJrKrLrMrPrQr r#r"r$rrrrNc stjst|}t|D]\}}| s0||n|| }||}||||}tjs| rt}|jj |jj kr|jj |vsJd|dt |rt |}t |}t |}t |d7}| r|nt |}|dkr | r| d||n|j||d}||d| |j||d|dd||}d||||}dd|dd|||fdd }fd d }| rtd k||d }|j|||ddqd kr|j||||ddq|j||ddqdS)NIIf capturable=True, params and state_steps must be on supported devices: .rralpha)valuecs,dddddS)Nr^?r0r0)rho_infrho_tr0r1 _compute_rectDsz+_single_tensor_radam.._compute_rectcs.}r|}n |}d|S)Nr`)sqrtaddadd_)Zexp_avg_sq_sqrt)bias_correction2rr"rCr0r1_compute_adaptive_lrLs   z2_single_tensor_radam.._compute_adaptive_lr@r&g)r=jit is_scriptingr enumeratecompiler is_compilingr r5typerEZ view_as_realrZmul_reZlerp_Zaddcmul_whererf)rrJrKrLrMrPrQr r#r"r$rrrrNiparamrDrBZstep_tcapturable_supported_devicesr2bias_correction1Zbias_corrected_exp_avgrcrhupdater0)rgrr"rCrarbr1_single_tensor_radamsh             rvc %sDt|dkrdS| rJdtjs`| r`tddtfddt||Ds`Jddtt |||||g}| D]\\}}}}}}t t t |}t t t |}t t t |}t t t |}t t t |}tjs|djrtj|tjd d d d d n t|d |r2t||||| rBt|}dd d | rt|}t|t|d t|}t||t|dt||t|t||}nfdd|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||~nffdd|D}!dd|!D}#fdd|D}fddt|#|D}"fddt||!|D}t|}$t|$| t|$|t|$t|$|"t|||$qdS)Nrz#_foreach ops don't support autogradF)Z supports_xlac3s.|]&\}}|jj|jjko$|jjvVqdSN)r5ro).0rAr2)rsr0r1 sz&_multi_tensor_radam..rYrZr&cpu)r5r[rr^cs8g|]0}dt|t|dt|qS)r^rrrxr2)rQrar0r1 s z'_multi_tensor_radam..r_cSs"g|]\}}t|dk|dqS)rir%r=rp)rxnrbr0r0r1r}scSsg|]}t|dkddqS)rr%r&r~rxrectr0r0r1r}csDg|]<}|dkr<|d|ddd|dndqS)r_r^r`rr0)rxrb)rar0r1r} s cSsg|]}|dkrdndqS)rr&r0rr0r0r1r}rcsg|]}dt|qS)rr{r|)rPr0r1r}scs g|]\}}||dqS)r0)rxrbc)r r0r1r}scs6g|].\}}}dt|d||dqS)rr`rr{)rxr2rr)rQr r0r1r} s)r<r=rmrnr allziprrZ"_group_tensors_by_device_and_dtypevaluesrlistrZis_cpuZ _foreach_add_r?rZ _foreach_negZ _foreach_powZ _foreach_neg_Z _foreach_mul_Z _foreach_div_Z _foreach_addZ_foreach_lerp_Z_foreach_addcmul_Z _foreach_subZ _foreach_mulZ_foreach_sqrt_Z _foreach_sqrtZ_foreach_reciprocal_)%rrJrKrLrMrPrQr r#r"r$rrrrNZgrouped_tensorsZgrouped_params_Zgrouped_grads_Zgrouped_exp_avgs_Zgrouped_exp_avg_sqs_Zgrouped_state_steps__Zgrouped_paramsZ grouped_gradsZgrouped_exp_avgsZgrouped_exp_avg_sqsZgrouped_state_stepsrtrgZ rho_t_listnumZsub2ZdenomrZunrect_step_sizeZ unrectifiedbufferr0)rPrQrsr rar1_multi_tensor_radamhs                                                       r)Zsingle_tensor_fnF)rrJrKrLrMr$rrrrNrrPrQr r#r"c Cstdd|Dstd|dur4t||dd\}}|rJtjrJtd|r^tjs^t}nt}||||||| | | ||| |||| ddS) zpFunctional API that performs RAdam algorithm computation. See :class:`~torch.optim.RAdam` for details. css|]}t|tjVqdSrw)r'r=r)rxtr0r0r1ryIrzradam..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF)Z use_fusedz6torch.jit.script not supported with foreach optimizers) rPrQr r#r"rr$rrrN)rrGrr=rjrkrrv)rrJrKrLrMr$rrrrNrrPrQr r#r"rfuncr0r0r1r/s< )FNFFFF)__doc__typingrrrr=rZ optimizerrrr r r r r rrrrrrrr__all__rrr>rWrvrrr0r0r0r1s D3  Q i H