˲j%E ddlmZmZmZmZmZddlZddlmZddlm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZddgZGddeZd d ed e d ed e d e d ze_deedeedeedeedeedededededededededefdZdeedeedeedeedeedededededededededefdZe e d"deedeedeedeedeed eedededededededededefd!ZdS)#)castListOptionalTupleUnionN)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_use_grad_for_differentiable _view_as_real OptimizerParamsTAdamaxadamaxceZdZ dddddded eeefd eeefd ed ed ee de de de ffdZ fdZ dZ e ddZxZS)rMb`?g?g+?:0yE>rNF)maximizedifferentiable capturableparamslrbetaseps weight_decayforeachrrr c 6t|tr'|dkrtdd|kstd|d|kstd|d|dcxkrdksntd|dd|dcxkrdksntd |dd|kstd |t |||||||| } t || dS) Nr zTensor 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$r%r&rrr ) isinstancernumel ValueErrordictsuper__init__) selfr!r"r#r$r%r&rrr defaults __class__s CC:\Projects\Historical\venv\Lib\site-packages\torch/optim/adamax.pyr/zAdamax.__init__sb b& ! ! grouppp_statestep_valr2s r3r;zAdamax.__setstate__DsO U###&  E   Y - - -   Z / / /   -u 5 5 5   \5 1 1 18_  *..B//w<<1$$U_WV_-M-M$$WV_55H !.O $,=,?,?#\(:K:M:MNNN FO    r4cd}|dD]x}|j |tj|z}|||jjrt d||j|j|} t| dkr|dr(tjdt|j n!tj dt | d <tj |tj | d <tj |tj | d <|| d || d || d z|S)NFr!z(Adamax does not support sparse gradientsrr r7r(r:r6) memory_formatexp_avgexp_inf)gradrA is_complexappend is_sparse RuntimeErrorr>r@zerosrr9rD zeros_likepreserve_format) r0rEparams_with_gradgradsexp_avgsexp_infs state_steps has_complexrFr>s r3 _init_groupzAdamax._init_groupWs x . .Av~ 5+A.. .K  # #A & & &v O"#MNNN LL JqME5zzQ\*FEK*;*=*=ahOOOOc1B1D1DEEEf $)#3U%:$$$i $)#3U%:$$$i  OOE), - - - OOE), - - -   uV} - - - -r4c|d}|5tj5|}dddn #1swxYwY|jD]}g}g}g}g}g}|d\} } |d} |d} |d} |d}|d}|d}|d }|||||||}t |||||| | | | | ||||| |S) zPerforms a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr#r$r"r%r&rrr ) r$beta1beta2r"r%r&rrr r[) _cuda_graph_capture_health_checkrA enable_gradr<r\r)r0closurelossrErVrWrXrYrZr^r_r$r"r%r&rrr r[s r3r6z Adamax.stepzs --///  "$$ ! !wyy ! ! ! ! ! ! ! ! ! ! ! ! ! ! !&$ $ E-/ "$E%'H%'H(*K >LE5,CtB 0LI&GZ(H"#34N|,J**'(KK  )!-%'     $ s AA A)rrrrNN)__name__ __module__ __qualname__rrrCrrrboolr/r;r\rr6 __classcell__)r2s@r3rrs!$(%1"&$+$ $+$+$+$+ %- $+UE\" $+  $+  $+$$+$+$+$+$+$+$+$+$+L&!!!F"444"!44444r4aImplements Adamax algorithm (a variant of Adam based on infinity norm). .. 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{ (weight decay)}, \\ &\hspace{13mm} \epsilon \text{ (epsilon)} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, u_0 \leftarrow 0 \text{ ( infinity norm)} \\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{5mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm}if \: \lambda \neq 0 \\ &\hspace{10mm} g_t \leftarrow g_t + \lambda \theta_{t-1} \\ &\hspace{5mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{5mm}u_t \leftarrow \mathrm{max}(\beta_2 u_{t-1}, |g_{t}|+\epsilon) \\ &\hspace{5mm}\theta_t \leftarrow \theta_{t-1} - \frac{\gamma m_t}{(1-\beta^t_1) u_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 `Adam: A Method for Stochastic Optimization`_. z Args: a lr (float, Tensor, optional): learning rate (default: 2e-3) betas (Tuple[float, float], optional): coefficients used for computing running averages of gradient and its square 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) z zd .. _Adam\: A Method for Stochastic Optimization: https://arxiv.org/abs/1412.6980 r!rWrXrYrZr$r^r_r"r%rrr r[c Ht|D]\}}||}| s|n| }||}||}||}tjsF| rDt }|jj|jjkr|jj|vsJd|d|dz }| dkr||| }tj|rPtj |}tj |}tj |}tj |}| |d|z | sPtj | || ||ntj| |d| |dgd}|tj|dd| r:||zdz }||||z}|||]d|t-|zz }||z }|||| dS) NIIf capturable=True, params and state_steps must be on supported devices: .r ralpha)outF)keepdim)value) enumeraterAcompiler is_compilingrr9typeaddrO view_as_reallerp_maximummul_absadd_cat unsqueeze unsqueeze_copy_amaxdiv_addcdiv_r)r!rWrXrYrZr$r^r_r"r%rrr r[iparamrNrLrMstep_tcapturable_supported_devicesnorm_bufneg_bias_correctiondenombias_correctionclrs r3_single_tensor_adamaxrs"f%%69695Qx#.tt$1+1+Q~**,, { {+L+N+N ( !V]%777L%)EEEEz[wzzzFEF !  1  88E866D  E " " 2&u--E%d++D(11G(11G  dAI&&& B M U## $$      ye$$..q11488::??33G3G3R3RST3U3UVH MM%*Xq%@@@ A A A  9#(-!"3   $ $R ( ( (11E NN7E * * * *%:f+=+="==O&C NN7GC4N 8 8 8 8m6969r4c | r Jdt|dkrdStjsI| rGt dt fdt ||DsJddtj|||||g}| D]\\}}}}}}ttt|}ttt|}ttt|}ttt|}ttt|}| rt||||| rtj|}tjs9|djr,tj|tjdd d ntj|d | dkr1| rtj||| ntj||| }tj||d z tj||| s| dkrtj|}ntj|tj||tj||| rltj|}tj|d tj|tj||}tj|||pfd |D}fd|D}tj||||dS)Nz#_foreach ops don't support autogradrF) supports_xlac3nK|]/\}}|jj|jjko |jjvV0dSrd)r9ru).0rFr6rs r3 z'_multi_tensor_adamax..Gs]  4 HMT[- - > !==      r4rkrlr)cpu)r9rmr c:g|]}dt|zz S)r r)rr6r^s r3 z(_multi_tensor_adamax..s8   26EZ----   r4c:g|]}t|z dzS)r)rbcr"s r3rz(_multi_tensor_adamax..s)OOO*R..2-3OOOr4)r@rArsrtrallzipr"_group_tensors_by_device_and_dtypevaluesrrrr _foreach_negis_cpu _foreach_add_rD _foreach_add_foreach_lerp_ _foreach_mul_ _foreach_abs _foreach_abs__foreach_maximum_ _foreach_pow _foreach_sub_ _foreach_div_ _foreach_mul_foreach_addcdiv_)r!rWrXrYrZr$r^r_r"r%rrr r[grouped_tensorsgrouped_params_grouped_grads_grouped_exp_avgs_grouped_exp_infs_grouped_state_steps__grouped_params grouped_gradsgrouped_exp_avgsgrouped_exp_infsgrouped_state_stepsbias_correctionsr step_sizers ` ` @r3_multi_tensor_adamaxr,s"DDDDD  6{{a > & & ( (wZw'H( ( ( $    v{33      w w wWs v v v  w w  B (K8O  " " $ $ II  d6lO<<T&\>:: V .?@@V .?@@"4<1EFF    /?AQ     >!.}==M ~**,, 81DQ1G1N 8  #U\#e%D%D%DC        3Q 7 7 7 1   #M>VVVVV % 2!>!!! -}a%iHHH ,e444 /LA--!.}==MM   . . . M3///  0-@@@  $1%9LMM    0! 4 4 4   0" 5 5 5&'79IJJE  #N4De L L L L    :M    POOO>NOOOI  # 02BI    OIIr4)single_tensor_fnFr&c tjs(td|Dst d|t ||d\}}|r-tjrt d|r&tjst}nt}||||||| | | | |||| |dS)zrFunctional API that performs adamax algorithm computation. See :class:`~torch.optim.Adamax` for details. c3JK|]}t|tjVdSrd)r*rAr)rts r3rzadamax..s?55() 1el##555555r4zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF) use_fusedz6torch.jit.script not supported with foreach optimizers) r$r^r_r"r%rrr[r ) rArsrtrrRr jit is_scriptingrr)r!rWrXrYrZr&rrr r[r$r^r_r"r%rfuncs r3rrs*4 > & & ( ( 55-855522  ^   1 Ne   7U59))++USTTT%uy--//%#$D   !%r4)NFFFF)typingrrrrrrAr optimizerr r r r rrrrrrrrrr__all__r__doc__rCrhrrrrJr4r3rs65555555555555 $ X RRRRRYRRRl4     5+`G9 LG9 <G96lG96l G9 f G9 G9 G9 G9 G9G9G9G9G9G9G9G9G9Tm Lm <m6lm6l m f m m m m mmmmmmmmm` 1FGGG# << L< <<6l<6l < f <d^<<<<< < < !<" #<$%<<<HG<<<r4