˲ja"dZddlmZmZmZmZmZddlZddlmZddl m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZddgZGd deZd d ed ed ed e d e d ze_deedeedeedeedeededededededededededefdZdeedeedeedeedeededededededededededefdZe e  d$deedeedeedeedeeded"eedededededededededef d#ZdS)%z'Implementation for the RAdam algorithm.)castListOptionalTupleUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc _params_doc_use_grad_for_differentiable _view_as_real OptimizerParamsTRAdamradamceZdZ ddddddded eeefd eeefd ed ed ede edededeffdZ fdZ dZ e ddZxZS)rMbP?g?g+?:0yE>rFN)foreachmaximize capturabledifferentiableparamslrbetaseps weight_decaydecoupled_weight_decayrrr r!c 8t|tr'|dkrtdd|kstd|d|kstd|d|dcxkrdksntd|dd|dcxkrdksntd |dd|kstd |t ||||||| || } t || dS) Nr zTensor lr must be 1-elementzInvalid learning rate: zInvalid epsilon value: r?z#Invalid beta parameter at index 0: z#Invalid beta parameter at index 1: zInvalid weight_decay value: ) r#r$r%r&rrr r'r!) isinstancernumel ValueErrordictsuper__init__) selfr"r#r$r%r&r'rrr r!defaults __class__s BC:\Projects\Historical\venv\Lib\site-packages\torch/optim/radam.pyr0zRAdam.__init__se b& ! ! (((E)$+0*:)>+++E,'i 0111""5#6777""5=111r5c|d}|5tj5|}dddn #1swxYwY|jD]}g}g}g}g}g}t t t t f|d\} } |||||||} t|||||| | |d|d|d|d|d|d|d |d | |S) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr$r#r&r%rrr r!r') beta1beta2r#r&r%rrr r!r'r\) _cuda_graph_capture_health_checkrB enable_gradr=rrrDr]r) r1closurelossrFrWrXrYrZr[r_r`r\s r4r7z RAdam.step~s~ --///  "$$ ! !wyy ! ! ! ! ! ! ! ! ! ! ! ! ! ! !&  E-/ "$E%'H(*K(*KeUl 3U7^DDLE5**'+{K  ;">2%Lz*i( .$%56',-E'F'!     & s AA A)rrrrFN)__name__ __module__ __qualname__rrrDrrboolrr0r<r]rr7 __classcell__)r3s@r4rrs.$(%1',&+#' $&+&+&+&+ %- &+UE\" &+  &+  &+!%&+$&+&+&+&+&+&+&+&+&+P(!!!F"---"!-----r5aImplements 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: aD 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 use decoupled weight decay as in AdamW to obtain RAdamW (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 r"rXrYrZr[r_r`r#r&r%r'r!rr r\c  t|D]r\}}| s||n|| }||}||||}tjsF| rDt }|jj|jjkr|jj|vsJd|dtj|rPtj|}tj|}tj|}tj|dz }| r|nt|}|dkr5| r| d||zz n| ||}| |d|z  | ||d|z d||zz }d||zz ||z }dd|z z dz d|z||zzz z fd} fd }| rLtjd k||zd }|||z|zd  d kr2|||z|z|zd X|||zd tdS) NIIf capturable=True, params and state_steps must be on supported devices: .r ralpha)valuecDdz dz zzdz dz zzz dzS)Nrq?rK)rho_infrho_tsr4 _compute_rectz+_single_tensor_radam.._compute_rect=sI19aKGaK058:  r5c}r|}n|}dz|z S)Nrt)sqrtaddadd_)exp_avg_sq_sqrtbias_correction2r!r%rNs r4_compute_adaptive_lrz2_single_tensor_radam.._compute_adaptive_lrEsR(oo//O <"1"5"5c":":"1"6"6s";";$c)_< z&_multi_tensor_radam..}s]  4 HMT[- - > !==      r5rlrmr*cpu)r:rnr rqc g|]@}dt|zt|zzdt|zz z z AS)rqr r)rr7r`rus r4 z'_multi_tensor_radam..st  T""#Jt,,,.u 4 0 00022r5rscHg|]\}}tj|dk|d S)rr)rBr)rnrvs r4rz'_multi_tensor_radam..s;5=Q ECKC00r5cBg|]}tj|dkddS)rr)r*rrrects r4rz'_multi_tensor_radam..s*QQQD D1Hc3 ? ?QQQr5c`g|]*}|dkr |dz |dz zzdz dz z|zz dznd+S)rsrqrtrrK)rrvrus r4rz'_multi_tensor_radam..s~   199 QYqy" ! ! 4u<>     r5c"g|] }|dkrdnd S)rr*rKrs r4rz'_multi_tensor_radam.. s$CCCdq11cCCCr5c:g|]}dt|zz S)r r)rr7r_s r4rz'_multi_tensor_radam..s8   26EZ----   r5c,g|]\}}|z|z dzS)rK)rrbcr#s r4rz'_multi_tensor_radam..s7   *2$dR2%   r5c`g|]*\}}}dt|zz dz|z|z zdz+S)r rtrr)rr7rrr`r#s r4rz'_multi_tensor_radam..sX   "D$ez$////C7BINKbP   r5)rArBrrrallzipr"_group_tensors_by_device_and_dtypevaluesrrris_cpu _foreach_add_rEr _foreach_neg _foreach_pow _foreach_neg_ _foreach_mul_ _foreach_div_ _foreach_add_foreach_lerp__foreach_addcmul_ _foreach_sub _foreach_mul_foreach_sqrt_ _foreach_sqrt_foreach_reciprocal_)'r"rXrYrZr[r_r`r#r&r%r'r!rr r\grouped_tensorsgrouped_params_grouped_grads_grouped_exp_avgs_grouped_exp_avg_sqs_grouped_state_steps__grouped_params grouped_gradsgrouped_exp_avgsgrouped_exp_avg_sqsgrouped_state_stepsrr} rho_t_listnumsub2denomrunrect_step_size unrectifiedbufferrrus' ``` @@r4_multi_tensor_radamras$ 6{{aDDDDD  > & & ( (wZw'H( ( ( $    v{33      w w wWs v v v  w w  B +{;O  " " $ $ [J[J  d6lO<<T&\>:: V .?@@"4<1EFF"4<1EFF ~**,, 81DQ1G1N 8  #U\#e%D%D%DC        3Q 7 7 7    /?AT     >!.}==Mq5y/A%  $1%9LMM    0 1 1 1   0! 4 4 4$1%9LMM    02E F F F   0! 4 4 4   02B C C C   0 1 1 1   0' : : :)JJ 0 J 1  % #NA\8I4IJJJJ'%~\%*$6%~\%%%M -}a%iHHH /777   q5y    > $Z33C%j!44D  T * * *  W - - -{w{3G&z7;;E  U + + +   % % %ADS*AUAUDQQDQQQ    0" 5 5 5$1%9LMM    0 1 1 1   0! 4 4 4   02B C C C   0 1 1 1$1%9LMM    0 1 1 1   0! 4 4 4  !1 2 2 2   0" 5 5 5   0$ 7 7 7   0 1 1 1   02B C C C      (   DDCdCCCK    :M        69+GW6X6X         &)*=tEU&V&V     $%899 FC((( F$4555 "6*** F$4555 0@&IIIIw[J[Jr5)single_tensor_fnFrc ttd|Dstd|t||d\}}|r-tjrtd|r&tjst }nt}||||||| | | ||| |||| dS)zpFunctional API that performs RAdam algorithm computation. See :class:`~torch.optim.RAdam` for details. c3JK|]}t|tjVdSre)r+rBr)rts r4rzradam..>s.@@qz!U\**@@@@@@r5zPAPI 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%rr'r!r r\)rrSr rBjit is_scriptingrr)r"rXrYrZr[r'rr!r r\rr_r`r#r&r%rfuncs r4rr$s4 @@K@@@ @ @  ^   1 Ne   7U59))++USTTT$uy--//$"#D  ! 5%r5)FNFFFF)__doc__typingrrrrrrBr optimizerr r r r rrrrrrrrrr__all__rrDrirrrrKr5r4rsc/.55555555555555 $ G NNNNNINNNd2f     gJ ^`D L`D <`D6l`Df `D f `D `D `D `D`D `D!`D`D`D`D !`D`D`D`DF@J L@J <@J6l@Jf @J f @J @J @J @J@J @J!@J@J@J@J !@J@J@J@JF 1EFFF$)" ;; L; <;6l;f ; f ;!;d^;;;;; ; !;" #;$%;& ';;;GF;;;r5