U MhRA@snddlmZmZmZmZddlZddlmZddlmZm Z m Z m Z m Z m Z mZmZmZmZmZmZmZddgZGdddeZd d e d ed e d ed e_eeeeeeeeeeeeeeeeeeed ddZeeeeeeeeeeeeeeeeeeed ddZe eddeeeeeeeeeeeeeeeeeeeeedddZdS))ListOptionalTupleUnionN)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_use_grad_for_differentiable _view_as_real OptimizerParamsTAdamaxadamaxc sleZdZdddddeeeeefeeeeeeed fd d Zfd d Z d dZ e dddZ Z S)rMb`?g?g+?:0yE>rNF)maximizedifferentiable capturable) paramslrbetaseps weight_decayforeachrrrc sd|kstd|d|ks,td|d|dkrDdksXntd|dd|dkrpdksntd|dd|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: )rrr r!r"rrr) ValueErrordictsuper__init__) selfrrrr r!r"rrrdefaults __class__D/var/www/html/venv/lib/python3.8/site-packages/torch/optim/adamax.pyr(s*  zAdamax.__init__cst||jD]}|dd|dd|dd|dd|dD]h}|j|g}t|dkrNt|dsNt |d}|drtj |t |j d ntj |t d |d<qNqdS) Nr"rFrrrrstepdtypedevicer1) r' __setstate__ param_groups setdefaultstategetlentorchZ is_tensorfloattensorrr2)r)r7grouppZp_stateZstep_valr+r-r.r4?s$        zAdamax.__setstate__c Csd}|dD]}|jdkrq |t|O}|||jjrDtd||j|j|} t| dkr|drtjdt |j dntj dt d | d <tj |tj d | d <tj |tj d | d <|| d || d || d q |S)NFrz(Adamax does not support sparse gradientsrrr-r0r#r3r/)Z memory_formatexp_avgexp_inf)gradr: is_complexappendZ is_sparse RuntimeErrorr7r9Zzerosrr2r<Z zeros_likeZpreserve_format) r)r=params_with_gradgradsexp_avgsexp_infs state_steps has_complexr>r7r-r-r. _init_groupRs6        zAdamax._init_groupcCs|d}|dk r.t |}W5QRX|jD]}g}g}g}g}g}|d\} } |d} |d} |d} |d}|d}|d}|d }|||||||}t|||||| | | | | |||||d q4|S) zPerforms a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nrr rr!r"rrr) r beta1beta2rr!r"rrrrJ)Z _cuda_graph_capture_health_checkr:Z enable_gradr5rKr)r)closureZlossr=rErFrGrHrIrLrMr rr!r"rrrrJr-r-r.r/usZ   z Adamax.step)rrrrN)N)__name__ __module__ __qualname__rr;rrboolr(r4rKrr/ __classcell__r-r-r+r.rs0  $ #aImplements 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`_. 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 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 )rrFrGrHrIr rLrMrr!rrrrJc Cst|D]\}}||}| s"|n| }||}||}||}tjs| rt}|jj|jjkrp|jj|kstd|d|d7}| dkr|j|| d}t |rt |}t |}t |}t |}| |d|| stj | ||||dnBt| |d||dgd}|tj|ddd| r|||d}||||}|||qd|t|}||}|j||| d qdS) NIIf capturable=True, params and state_steps must be on supported devices: .rralpha)outF)Zkeepdim)value) enumerater:_utils is_compilingr r2typeAssertionErroraddrBZ view_as_realZlerp_maximumZmul_absZadd_catZ unsqueezeZ unsqueeze_Zcopy_ZamaxZdiv_Zaddcdiv_r)rrFrGrHrIr rLrMrr!rrrrJiparamrAr?r@Zstep_tcapturable_supported_devicesZnorm_bufZneg_bias_correctiondenomZbias_correctionZclrr-r-r._single_tensor_adamaxsT        "  rgc s| r tdt|dkrdStjs`| r`tddtfddt||Ds`tddt |||||g}| D]l\\}}}}}}| rt ||||| rt |}|dj rtj|tjd d d d d n t|d | dkr| rtj||| d ntj||| d }t||d t||| sN| dkrNt|}n t|t||t||| rt|}t|d t|t||}t|||q|fdd|D}fdd|D}t||||q|dS)Nz#_foreach ops don't support autogradrF)Z supports_xlac3s.|]&\}}|jj|jjko$|jjkVqdSN)r2r]).0r>r/)rer-r. Csz'_multi_tensor_adamax..rTrUr$cpu)r2rVrcsg|]}dt|qS)rr)rir/)rLr-r. sz(_multi_tensor_adamax..csg|]}t|dqS)rl)ribc)rr-r.rms)r^r9r:r[r\r allziprZ"_group_tensors_by_device_and_dtypevaluesrZ _foreach_negZis_cpuZ _foreach_add_r<Z _foreach_addZ_foreach_lerp_Z _foreach_mul_Z _foreach_absZ _foreach_abs_Z_foreach_maximum_Z _foreach_powZ _foreach_sub_Z _foreach_div_Z _foreach_mulZ_foreach_addcdiv_)rrFrGrHrIr rLrMrr!rrrrJZgrouped_tensorsZgrouped_paramsZ grouped_gradsZgrouped_exp_avgsZgrouped_exp_infsZgrouped_state_steps_Zbias_correctionsrfZ step_sizer-)rLrerr._multi_tensor_adamax(s                       rt)Zsingle_tensor_fnF)rrFrGrHrIr"rrrrJr rLrMrr!c Cstjs$tdd|Ds$td|dkr>t||dd\}}|rTtjrTtd|rhtjsht}nt }||||||| | | | |||| |ddS) zrFunctional API that performs adamax algorithm computation. See :class:`~torch.optim.Adamax` for details. css|]}t|tjVqdSrh) isinstancer:r)ritr-r-r.rjszadamax..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF)Z use_fusedz6torch.jit.script not supported with foreach optimizers) r rLrMrr!rrrJr) r:r[r\rprDr ZjitZ is_scriptingrtrg)rrFrGrHrIr"rrrrJr rLrMrr!rsfuncr-r-r.rsB )NFFFF)typingrrrrr:rZ optimizerrr r r r r rrrrrrr__all__r__doc__r;rRrgrtrr-r-r-r.s <    2 K j