o <&iq)@sddlZddlmZddlmZmZmZmZmZmZm Z m Z m Z m Z m Z mZmZddlmZmZmZmZddlmZddgZGd ddeZd d e d e d ed e d e d e_       d+deedeedeedeedeedeedeedededeedeedeedededededeeefd ed!ed"ef(d#dZdeedeedeedeedeedeedeedeededededeeefd ed!ed"edededef$d$d%Zdeedeedeedeedeedeedeedeededededeeefd ed!ed"edededef$d&d'Zdeedeedeedeedeedeedeedeededededeeefd ed!ed"edededed(df&d)d*ZdS),N)Tensor) Optimizer_use_grad_for_differentiable _get_value_dispatch_sqrt_stack_if_compiling_capturable_doc_differentiable_doc _foreach_doc _fused_doc _maximize_doc_default_to_fused_or_foreachParamsT _view_as_real)ListOptionalTupleUnion)$_get_fused_kernels_supported_devicesAdamWadamwcseZdZ     dddddddded eeefd eeefd ed ed edede ededede effddZ fddZ ddZ e dddZZS)rMbP?g?g+?:0yE>{Gz?FN)maximizeforeach capturabledifferentiablefusedparamslrbetaseps weight_decayamsgradrrrrr c s6d|ks td|t|tr|r| stdd|ks#td|d|dkr/dks9ntd|dd|dkrEdksOntd |dd|ksZtd |t|||||||| | | d } t|| | r| rwtd d |_tt fdd|j Dstdd|rtddSdS)NzInvalid learning rate: Elr as a Tensor is not supported for capturable=False and foreach=TruezInvalid epsilon value: r?z#Invalid beta parameter at index 0: rz#Invalid beta parameter at index 1: zInvalid weight_decay value: ) r"r#r$r%r&rrrrr z)`fused` does not support `differentiable`Tc3s4|]}|dD]}|jjvot|VqqdS)r!N)devicetypetorchZis_floating_point).0ZpgpZfused_supported_devices?s  z!AdamW.__init__..zX`fused=True` requires all the params to be floating point Tensors of supported devices: .z0`fused` and `foreach` cannot be `True` together.) ValueError isinstancerdictsuper__init__ RuntimeErrorZ_step_supports_amp_scalingrall param_groups) selfr!r"r#r$r%r&rrrrr defaults __class__r/r1r8 sP   zAdamW.__init__cst||jD]&}|dd|dd|dd|dd|dd|ddq t|j}t|dkoEt |dd }|s[|D]}tj t |d tj d |d <qJdSdS) Nr&Frrrrr rstepdtype) r7 __setstate__r; setdefaultliststatevalueslenr,Z is_tensortensorfloatfloat32)r<rFgroupZ state_valuesZstep_is_tensorsr>r0r1rCIs"        zAdamW.__setstate__c Csdd} |dD]} | jdurq| t| O} || | jjr"td|| j|j| } t| dkro|ds;|drEtjdtj | j dntj d tj d | d <tj | tj d | d <tj | tj d | d<|rotj | tj d | d<|| d || d|dr|| d|dr| d jrtd|drt|dtr|dstd|| d q| S)NFr!z'AdamW does not support sparse gradientsrrr r0)rBr*r'rAr@)Z memory_formatexp_avg exp_avg_sqmax_exp_avg_sqr&rzB`requires_grad` is not supported for `step` in differentiable moderr"r()gradr, is_complexappendZ is_sparser9rFrHZzerosrKr*rIZ zeros_likeZpreserve_formatZ requires_gradr5r) r<rLparams_with_gradgradsr&exp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexr.rFr0r0r1 _init_groupZsJ         zAdamW._init_groupcCs|d}|dur!t |}Wdn1swY|jD]V}g}g}g}g}g}g} |d} |d\} } ||||| |||| } t|||||| f| | | |d|d|d|d|d|d |d |d t|d dt|d d| dq$|S)zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr&r#r"r%r$rrrrr grad_scale found_inf)r&beta1beta2r"r%r$rrrrr r\r]rZ)Z _cuda_graph_capture_health_checkr,Z enable_gradr;r[rgetattr)r<closureZlossrLrTrUrVrWrXrYr&r^r_rZr0r0r1r@sb       z AdamW.step)rrrrFN)__name__ __module__ __qualname__rrrJrrboolrr8rCr[rr@ __classcell__r0r0r>r1r sN         <=a Implements AdamW 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)}, \: \epsilon \text{ (epsilon)} \\ &\hspace{13mm} \lambda \text{(weight decay)}, \: \textit{amsgrad}, \: \textit{maximize} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ (first moment)}, v_0 \leftarrow 0 \text{ ( second moment)}, \: \widehat{v_0}^{max}\leftarrow 0 \\[-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} - \gamma \lambda \theta_{t-1} \\ &\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 m_t/\big(1-\beta_1^t \big) \\ &\hspace{5mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\textbf{if} \: amsgrad \\ &\hspace{10mm}\widehat{v_t}^{max} \leftarrow \mathrm{max}(\widehat{v_t}^{max}, \widehat{v_t}) \\ &\hspace{10mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}^{max}} + \epsilon \big) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}\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 `Decoupled Weight Decay Regularization`_. a Args: params (iterable): iterable of parameters to optimize or dicts defining parameter groups lr (float, Tensor, optional): learning rate (default: 1e-3). A tensor LR is not yet supported for all our implementations. Please use a float LR if you are not also specifying fused=True or capturable=True. 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 coefficient (default: 1e-2) amsgrad (bool, optional): whether to use the AMSGrad variant of this algorithm from the paper `On the Convergence of Adam and Beyond`_ (default: False) z z .. _Decoupled Weight Decay Regularization: https://arxiv.org/abs/1711.05101 .. _On the Convergence of Adam and Beyond: https://openreview.net/forum?id=ryQu7f-RZ Fr!rUrVrWrXrYrrrr r\r]rZr&r^r_r"r%r$rc Cstjstdd|Dstd| dur.|dur.t||dd\}}|r.t|tr.|s.d}| dur4d} |dur:d}|rEtj rEtd| rPtj rPtd| rZtj sZt }n |rdtj sdt }nt }|||||||| ||||||||| | | d dS) zpFunctional API that performs AdamW algorithm computation. See :class:`~torch.optim.AdamW` for details. css|] }t|tjVqdSrb)r5r,r)r-tr0r0r1r24szadamw..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF)Z use_fusedz6torch.jit.script not supported with foreach optimizersz4torch.jit.script not supported with fused optimizers) r&r^r_r"r%r$rrrr\r]rZ) r,_utils is_compilingr:r9rr5rjit is_scripting _fused_adamw_multi_tensor_adamw_single_tensor_adamw)r!rUrVrWrXrYrrrr r\r]rZr&r^r_r"r%r$r_funcr0r0r1rsR c Cs|dur|dus Jtjrt| tsJt|D]-\}}|s%||n|| }||}||}||}tjsM|rM|jrC|jsM|j rI|j sMJdt |rqt |}t |}t |}|rlt ||||<t |}|d7}| d| | | |d| | | j||d| d|s|r|}d| |}d| |}| |}|}|}|r|r||}n||}||t||||||| |}n |||| |}|||nEt|}d| |}d| |}| |}t|}|r"tj|||||d|||| }n ||| }|j||| d|rHt ||rHt||||<qdS)NzGIf capturable=True, params and state_steps must be CUDA or XLA tensors.r)value)out)r,rkrlr5rJ enumeraterirjis_cudaZis_xlarRZ view_as_realZmul_Zlerp_Zaddcmul_negsqrtcloneZcopy_maximumZadd_Zaddcdiv_rrZview_as_complex) r!rUrVrWrXrYr\r]r&r^r_r"r%r$rrrrZiparamrQrNrOZstep_tr@bias_correction1bias_correction2 step_sizeZ step_size_negbias_correction2_sqrtrPdenomr0r0r1rois~            roc st|dkrdSttr|stdtjs*|r*tddt||Ds*Jd|r0Jd|dur8|dus:Jt ||||||g}| D]\\}}}}}}}|r[t |}|ro|rht |||||nt |||||djrtj|tjddd dd nt|d | dkrt|d | t||d t|t|||d ~|rt|}t|}t|d t|d t|t|t|t||}|}|rt||t|}nt|}t||t|| t||t|||qIfd d |D}fdd |D}tfdd |D}dd |D}|rDt||t|}nt|}t||t|| t||||qIdS)Nrr(css |] \}}|jo |jVqdSrb)ru)r-r.r@r0r0r1r2s z&_multi_tensor_adamw..z@If capturable=True, params and state_steps must be CUDA tensors.z#_foreach ops don't support autogradr)cpu)r*)alpharcg|] }dt|qSrrr-r@)r^r0r1 Rz'_multi_tensor_adamw..crrrr)r_r0r1rSrcsg|]}|dqS)r0r-bc)r"r0r1rUscSsg|]}t|qSr0)rrr0r0r1rWs)rHr5rr9r,rirjr:zipr"_group_tensors_by_device_and_dtyperGZ _foreach_negrZis_cpu _foreach_add_rIZ _foreach_mul_Z_foreach_lerp_Z_foreach_addcmul_Z _foreach_pow _foreach_sub_Z _foreach_neg_Z _foreach_div_Z_foreach_reciprocal_Z_foreach_sqrt_Z_foreach_maximum_Z _foreach_sqrtZ_foreach_addcdiv_r)r!rUrVrWrXrYr\r]r&r^r_r"r%r$rrrrZgrouped_tensors device_params device_gradsdevice_exp_avgsdevice_exp_avg_sqsdevice_max_exp_avg_sqsdevice_state_stepsrpr|r}r~rZexp_avg_sq_sqrtr0)r^r_r"r1rns                            rnreturnc Cs~|sdS|r td|dur|j|ind}|dur|j|ind}t| tr1t| jdkr1| j| ind}t||||||g}|D]z\\}}\\}}}}}}}d\}}|duri||vre|j|dd||<||}|dur~||vrz|j|dd||<||}|dur||vr| j|dd||<||} t |dt j |||||||| | | | | |||d|durt ||gt |qBdS) Nz9Adam with fused=True does not support differentiable=Truer)NNT) non_blocking)r*rr) r&r"r^r_r%r$rr\r])r9r*r5rstrrritemstor,rZ _fused_adamw_rrH) r!rUrVrWrXrYr\r]r&r^r_r"r%r$rrrrZZgrad_scale_dictZfound_inf_dictZlr_dictrr*rprrrrrrZdevice_grad_scaleZdevice_found_infr0r0r1rmgsf& rm)NFFNNNF) r,rZ optimizerrrrrrr r r r r rrrtypingrrrrZtorch.utils._foreach_utilsr__all__r__doc__rfrJrrornrmr0r0r0r1sn < J&K       R     v