o <&i3@sddlZddlmZddlmZmZmZmZmZmZm Z m Z m Z ddl m Z mZddgZGdddeZd d e d e d ed e_  d!de ede ede ede ede edeededededededededefddZde ede ede ede ede ededededededededefddZde ede ede ede ede ededededededededefdd ZdS)"N)Tensor) Optimizer_use_grad_for_differentiable _get_value_stack_if_compiling_default_to_fused_or_foreach_differentiable_doc _maximize_doc _foreach_doc _view_as_real)ListOptionalAdamaxadamaxcsdeZdZ     dddddeed ed effd d Zfd dZddZedddZ Z S)rMb`?g?g+?:0yE>rNF)maximizedifferentiableforeachrrc sd|ks td|d|kstd|d|dkr"dks,ntd|dd|dkr8dksBntd|dd|ksMtd |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: )lrbetaseps weight_decayrrr) ValueErrordictsuper__init__) selfparamsrrrrrrrdefaults __class__=C:\wamp64\www\opt\env\Lib\site-packages\torch/optim/adamax.pyr s(  zAdamax.__init__cst||jD]}|dd|dd|ddq t|j}t|dko3t |dd}|sI|D]}tj t |dtj d|d<q8dSdS)NrrFrrstepZdtype) r __setstate__ param_groups setdefaultliststatevalueslentorchZ is_tensortensorfloatfloat32)r!r.groupZ state_valuesZstep_is_tensorsr$r&r'r*/s     zAdamax.__setstate__c Csd}|dD]`}|jdurq|t|O}|||jjr"td||j|j|} t| dkrQtjdtj d| d<tj |tj d| d <tj |tj d| d <|| d || d || dq|S) NFr"z(Adamax does not support sparse gradientsrrr)r()Z memory_formatexp_avgexp_inf) gradr1 is_complexappendZ is_sparse RuntimeErrorr.r0r2r4Z zeros_likeZpreserve_format) r!r5params_with_gradgradsexp_avgsexp_infs state_steps has_complexpr.r&r&r' _init_group=s,        zAdamax._init_groupcCsd}|durt |}Wdn1swY|jD]F}g}g}g}g}g}|d\} } |d} |d} |d} |d}|d}|d}|||||||}t|||||| | | | | ||||d q |S) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nrrrrrrr) rbeta1beta2rrrrrrB)r1Z enable_gradr+rDr)r!closureZlossr5r=r>r?r@rArErFrrrrrrrBr&r&r'r(YsH   z Adamax.step)rrrrNN) __name__ __module__ __qualname__rboolr r*rDrr( __classcell__r&r&r$r'r s(    "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 Fr"r>r?r@rArrrrBrrErFrrc Cstdd|Ds td|durt||dd\}}|r%tjr%td|r/tjs/t}nt}||||||| | | | | |||d dS) zrFunctional API that performs adamax algorithm computation. See :class:`~torch.optim.Adamax` for details. css|] }t|tjVqdSrH) isinstancer1r).0tr&r&r' szadamax..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF)Z use_fusedz6torch.jit.script not supported with foreach optimizers)rrErFrrrrrB)allr<rr1ZjitZ is_scripting_multi_tensor_adamax_single_tensor_adamax)r"r>r?r@rArrrrBrrErFrr_funcr&r&r'rs4 cCs(t|D]\} }|| }| s|n| }|| }|| }|| }|d7}| dkr.|j|| d}t|rGt|}t|}t|}t|}||d|t||d| | dgd}| sqtj |dd|dn | tj |dddd|t|}||}|j||| dqdS)NrralphaFkeepdimout)rZ)value) enumerateaddr1r:Z view_as_realZlerp_catZmul_ unsqueezeabsadd_ unsqueeze_ZamaxZcopy_rZaddcdiv_)r"r>r?r@rArrErFrrrrrBiparamr9r7r8Zstep_tnorm_bufbias_correctionclrr&r&r'rTs2     $rTc s~| rJdt|dkrdSt|||||g} | D]\\}}}}}}| r,t|}| r5t|||||djrHtj|tj dddddnt|d|dkre| r]tj|||dntj |||d}t ||dt ||t ||D]&\}}t|d|| dgd}tj|dd||fd qyfd d |D}tfd d |D}t||||qdS) Nz#_foreach ops don't support autogradrrcpu)ZdevicerWrFrYcsg|] }dt|qS)r)r)rOr()rEr&r' _sz(_multi_tensor_adamax..csg|]}d|qS)r&)rOrg)rr&r'rj`s)r0rZ"_group_tensors_by_device_and_dtyper/r1Z _foreach_negr Zis_cpuZ _foreach_add_r2Z _foreach_addZ_foreach_lerp_Z _foreach_mul_zipr_r`rarbrcmaxnewlongrZ_foreach_addcdiv_)r"r>r?r@rArErFrrrrrrBZgrouped_tensorsZgrouped_paramsZ grouped_gradsZgrouped_exp_avgsZgrouped_exp_infsZgrouped_state_stepsrUr8r9rfZbias_correctionsrhr&)rErr'rS%s6       rS)NFFF)r1rZ optimizerrrrrrr r r r typingr r__all__r__doc__rLr3rrTrSr&r&r&r's ,   5     8      5