U Mhc@sddlmZmZmZmZmZddlZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZddgZGdddeZd d ed ed e d e d e_eeeeeeeeeeeeeeeeeeeeeeed ddZeeeeeeeeeeeeeeeeeeeeeeed ddZe eddeeeeeeeeeeeeeeeeeeeeeeeeedddZdS))castListOptionalTupleUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported_dispatch_sqrt _foreach_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc_stack_if_compiling_use_grad_for_differentiable _view_as_real OptimizerParamsTNAdamnadamcsreZdZddddddeeeeefeeeeeeeeed fd d Zfd d Z ddZ e dddZ Z S)rMb`?g?g+?:0yE>rMbp?FN)foreachmaximize capturabledifferentiable) paramslrbetaseps weight_decaymomentum_decaydecoupled_weight_decayrrr r!c sd|kstd|d|ks,td|d|dkrDdksXntd|dd|dkrpdksntd|dd|kstd |d|kstd |t||||||| || | d } t|| dS) NzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: rz#Invalid beta parameter at index 1: zInvalid weight_decay value: zInvalid momentum_decay value: ) r#r$r%r&r'r(rrr r!) ValueErrordictsuper__init__) selfr"r#r$r%r&r'r(rrr r!defaults __class__C/var/www/html/venv/lib/python3.8/site-packages/torch/optim/nadam.pyr.s2 zNAdam.__init__cst||jD]}|dd|dd|dd|dd|dd|dD]}|j|g}t|dkrZt|d st |d }|drtj |t |j d ntj |t d |d <t|d sZ|d }|drtj |t |j d ntj |t d |d <qZqdS) NrFrr r!r(r"rstepdtypedevicer7 mu_product) r- __setstate__ param_groups setdefaultstategetlentorchZ is_tensorfloattensorrr8)r/r>grouppZp_stateZstep_valZ mu_prod_valr1r3r4r;Es<          zNAdam.__setstate__c Cs,d}|dD]} | jdk r |t| O}|| | jjrDtd|| j|j| } t| dkr|drtjdt | j dntj dt d | d <|drtj dt | j dntj d t d | d <tj | tjd | d<tj | tjd | d<|| d|| d|| d || d q |S)NFr"z'NAdam does not support sparse gradientsrr r3r6r)r9r5r*r:)Z memory_formatexp_avg exp_avg_sq)gradrA is_complexappendZ is_sparse RuntimeErrorr>r@Zzerosrr8rCZonesZ zeros_likeZpreserve_format) r/rDparams_with_gradgradsexp_avgs exp_avg_sqs mu_products state_steps has_complexrEr>r3r3r4 _init_groupcs>        zNAdam._init_groupc Cs|d}|dk r.t |}W5QRX|jD]}g}g}g}g}g}g} ttttf|d\} } |||||||| } t|||||| | | |d|d|d|d|d|d|d |d |d | d q4|S) zPerforms a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr$r#r&r'r%rr(rr r!) beta1beta2r#r&r'r%rr(rr r!rR) Z _cuda_graph_capture_health_checkrAZ enable_gradr<rrrBrSr) r/closureZlossrDrLrMrNrOrPrQrTrUrRr3r3r4r5sV   z NAdam.step)rrrrrF)N)__name__ __module__ __qualname__rrBrboolrr.r;rSrr5 __classcell__r3r3r1r4rs8  ) 2a Implements NAdam algorithm. .. math:: \begin{aligned} &\rule{110mm}{0.4pt} \\ &\textbf{input} : \gamma_t \text{ (lr)}, \: \beta_1,\beta_2 \text{ (betas)}, \: \theta_0 \text{ (params)}, \: f(\theta) \text{ (objective)} \\ &\hspace{13mm} \: \lambda \text{ (weight decay)}, \:\psi \text{ (momentum decay)} \\ &\hspace{13mm} \: \textit{decoupled\_weight\_decay}, \:\textit{maximize} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0 \leftarrow 0 \text{ ( second moment)} \\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{5mm}\textbf{if} \: \textit{maximize}: \\ &\hspace{10mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm} \theta_t \leftarrow \theta_{t-1} \\ &\hspace{5mm} \textbf{if} \: \lambda \neq 0 \\ &\hspace{10mm}\textbf{if} \: \textit{decoupled\_weight\_decay} \\ &\hspace{15mm} \theta_t \leftarrow \theta_{t-1} - \gamma \lambda \theta_{t-1} \\ &\hspace{10mm}\textbf{else} \\ &\hspace{15mm} g_t \leftarrow g_t + \lambda \theta_{t-1} \\ &\hspace{5mm} \mu_t \leftarrow \beta_1 \big(1 - \frac{1}{2} 0.96^{t \psi} \big) \\ &\hspace{5mm} \mu_{t+1} \leftarrow \beta_1 \big(1 - \frac{1}{2} 0.96^{(t+1)\psi}\big)\\ &\hspace{5mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{5mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{5mm}\widehat{m_t} \leftarrow \mu_{t+1} m_t/(1-\prod_{i=1}^{t+1}\mu_i)\\[-1.ex] & \hspace{11mm} + (1-\mu_t) g_t /(1-\prod_{i=1}^{t} \mu_{i}) \\ &\hspace{5mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}} + \epsilon \big) \\ &\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 `Incorporating Nesterov Momentum into Adam`_. a Args: params (iterable): iterable of parameters to optimize or dicts defining parameter groups lr (float, optional): learning rate (default: 2e-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) momentum_decay (float, optional): momentum momentum_decay (default: 4e-3) decoupled_weight_decay (bool, optional): whether to use decoupled weight decay as in AdamW to obtain NAdamW (default: False) z z .. _Incorporating Nesterov Momentum into Adam: https://openreview.net/forum?id=OM0jvwB8jIp57ZJjtNEZ .. _Decoupled Weight Decay Regularization: https://arxiv.org/abs/1711.05101 )r"rMrNrOrPrQrTrUr#r&r'r%r(rr r!rRc Csbt|D]R\}}| s||n|| }||}||}||}||}t|rzt|}t|}t|}t|}tjs|rt}|jj|jjkr|jjkrnn |jj|kst d|d|d7}|r|}nt |}d||}| dkr&| r| d|| n|j || d}|ddd|| }|ddd|d| }||9}| |d|| |j||d|d ||}|s|r| | }||}|| d|d|}|| |d|}||||||qt ||}|| |j||| d|dt |d |j||| |d|d qdS) NzVIf capturable=True, params, mu_products and state_steps must be on supported devices: .rralphar*?Q?)value) enumeraterArIZ view_as_real_utils is_compilingrr8typeAssertionErrorrZmul_addZlerp_Zaddcmul_divsqrtZaddcdiv_Zadd_)r"rMrNrOrPrQrTrUr#r&r'r%r(rr r!rRiparamrHrFrGr:Zstep_tcapturable_supported_devicesr5Zbias_correction2mumu_nextdenomZmu_product_nextr3r3r4_single_tensor_nadamsp                rpc "st|dkrdS|rtdtjsb|rbtddtfddt|||Dsbtddt ||||||g}| D]\\}}}}}}}|rt ||||| rt |}|dj rtj|tjd d d d d n t|d | dkr6| rt|d | n(| r&tj||| d ntj||| d }t||d t|t|||d t|}|r t|}td|}t|dt|d t|t|td|}t|dt|d t|~t|}t|d t|t|n:fdd|D}fdd|D}fdd|D}t||t||t|| ~|r&t|d t|t|d }t|t|||}~t||}t|t|d t|||} ~t||}!t|!| |t||!|qtfddt||D}tfddt||D} t||||t|||| qdS)Nrz#_foreach ops don't support autogradF)Z supports_xlac3sD|]<\}}}|jj|jjko*|jjkno:|jjkVqdSN)r8re).0rEmpr5)rlr3r4 s$z&_multi_tensor_nadam..zWIf capturable=True, params, mu_products, and state_steps must be on supported devices: r\r*cpu)r8r]rr`gcs g|]}tdt|qS)r)r rrrr5)rUr3r4 sz'_multi_tensor_nadam..cs(g|] }dddt|qS)r*r_r`rrvrTr'r3r4rwscs,g|]$}dddt|dqS)r*r_r`rrxrvryr3r4rwscs0g|](\}}td|dt|dqSr*rx)rrr:rmr#r3r4rw scs0g|](\}}t|dt||dqSrzrx)rrr:rnr|r3r4rw&s)r@rfrArcrdrallziprZ"_group_tensors_by_device_and_dtypevaluesrZ _foreach_negZis_cpuZ _foreach_add_rCZ _foreach_mul_Z _foreach_addZ_foreach_lerp_Z_foreach_addcmul_Z _foreach_sqrtZ _foreach_mulZ _foreach_powZ _foreach_sub_Z _foreach_neg_Z_foreach_sqrt_Z _foreach_div_Z _foreach_subZ_foreach_addcdiv_r)"r"rMrNrOrPrQrTrUr#r&r'r%r(rr r!rRZgrouped_tensorsZgrouped_paramsZ grouped_gradsZgrouped_exp_avgsZgrouped_exp_avg_sqsZgrouped_mu_productsZgrouped_state_steps_Zexp_avg_sq_sqrtexponentZmusZmu_nextsZbias_correction_sqrtroZstep_size_gradsZstep_size_expavg numeratorr3)rTrUrlr#r'r4_multi_tensor_nadamps                                               r)Zsingle_tensor_fnF)r"rMrNrOrPrQr(rr r!rRrrTrUr#r&r'r%c Cstdd|Dstdtdd|Ds4td|dkrNt|| dd\}}|rdtjrdtd |rxtjsxt}nt}|||||||| | |||| |||| | d dS) zpFunctional API that performs NAdam algorithm computation. See :class:`~torch.optim.NAdam` for details. css|]}t|tjVqdSrq isinstancerArrrtr3r3r4rtVsznadam..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorscss|]}t|tjVqdSrqrrr3r3r4rt[szPAPI has changed, `mu_products` argument must contain a list of singleton tensorsNF)Z use_fusedz6torch.jit.script not supported with foreach optimizers) rTrUr#r&r'rr(r%r r!rR)r}rKr rAZjitZ is_scriptingrrp)r"rMrNrOrPrQr(rr r!rRrrTrUr#r&r'r%rfuncr3r3r4r9sL )FNFFFF) typingrrrrrrArZ optimizerr r r r r rrrrrrrrrr__all__r__doc__rBrZrprrr3r3r3r4s D6' D _ J