a hR@snddlmZmZmZddlZddlmZddlmZmZm Z m Z m Z m Z m Z mZmZmZmZmZmZmZmZmZmZmZmZmZddgZGdddeZd d ed e d ed ed e d e d e_eeeeeeeeeeeeeeeeeeee efee efee efe e eeeedddZ!eeeeeeeeeeeeeeeeeeee efee efee efe e eeeedddZ"eeeeeeeeeeeeeeeeeee e ee efe e eeeeddddZ#e e!ddeeeeeeeeeeeeeeeeeeeeeeeeee e ee efe e edddZ$dS))castOptionalUnionN)Tensor)_capturable_doc_default_to_fused_or_foreach_device_dtype_check_for_fused_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc _fused_doc!_get_capturable_supported_devices_get_scalar_dtype _get_value _maximize_doc _params_doc_stack_if_compiling _to_scalar_use_grad_for_differentiable _view_as_real DeviceDictDeviceDtypeDict OptimizerParamsTAdamadamcseZdZddddddddeeeefeeeefeeeffeeee eeeee eed fd d Z fd d Z d dZ e dddZZS)rMbP?g?g+?:0yE>rFN)foreachmaximize capturabledifferentiablefuseddecoupled_weight_decay) paramslrbetaseps weight_decayamsgradr r!r"r#r$r%c st|tr.|r| std|dkr.tdd|ksDtd|d|ksZtd|d|dkrrdksntd |dd|dkrdksntd |dd|kstd |t|dtrt|dtst|dtrt|dtstd t|dtrL| s2|r2td |ddkrLtdt|dtr| sp|rptd|ddkrtdt|||||||| | | | d } t|| | r| rtdd|_ |rtddS)NElr as a Tensor is not supported for capturable=False and foreach=TruerTensor 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: z0betas must be either both floats or both TensorszKbetas[0] as a Tensor is not supported for capturable=False and foreach=Truez!Tensor betas[0] must be 1-elementzKbetas[1] as a Tensor is not supported for capturable=False and foreach=Truez!Tensor betas[1] must be 1-element) r'r(r)r*r+r!r r"r#r$r%z)`fused` does not support `differentiable`Tz0`fused` and `foreach` cannot be `True` together.) isinstancer ValueErrornumelfloatdictsuper__init__ RuntimeErrorZ_step_supports_amp_scaling)selfr&r'r(r)r*r+r r!r"r#r$r%defaults __class__>/var/www/auris/lib/python3.9/site-packages/torch/optim/adam.pyr6#sx         z Adam.__init__cst||jD]}|dd|dd|dd|dd|dd|dd|dd}|d D]t}|j|g}t|d krrt|d srt |d }|ds|drtj |t |d |j d ntj |t d|d <qrqdS)Nr+Fr!r r"r#r%r$r&rstepZis_fuseddtypedevicerA) r5 __setstate__ param_groups setdefaultstategetlentorch is_tensorr3tensorrrB)r8rGgroupr$pZp_stateZstep_valr:r<r=rDrs.           zAdam.__setstate__c Csd}|dD]x} | jdur |t| O}|| | jjrDtd|| j|j| } t| dkr|drvt| |ds|drtj dt |dd| j d ntj d t d | d <tj | tjd | d<tj | tjd | d<|drtj | tjd | d<|| d|| d|dr.|| d|drL| d jrLtd|drxt|drx|dsxtd|| d q |S)NFr&zJAdam does not support sparse gradients, please consider SparseAdam insteadrr$r"r<r?r@r.rCr>)Z memory_formatexp_avg exp_avg_sqr+max_exp_avg_sqr#zB`requires_grad` is not supported for `step` in differentiable moder r'r,)gradrJ is_complexappendZ is_sparser7rGrIr zerosrrBrLZ zeros_likeZpreserve_format requires_gradrK) r8rMparams_with_gradgradsexp_avgs exp_avg_sqsmax_exp_avg_sqs state_steps has_complexrNrGr<r<r= _init_groupsj              zAdam._init_groupc Cs|d}|durBt|}Wdn1s80Y|jD]}g}g}g}g}g}g} |d\} } |||||||| } t|||||| |d| | | |d|d|d|d|d|d |d |d t|d dt|d d|ddqH|S)zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nr(r+r'r*r)r!r r"r#r$ grad_scale found_infr%)r+r]beta1beta2r'r*r)r!r r"r#r$r_r`r%)Z _cuda_graph_capture_health_checkrJZ enable_gradrEr^rgetattr) r8closureZlossrMrWrXrYrZr[r\rarbr]r<r<r=r>s\ $     z Adam.step)rrrrF)N)__name__ __module__ __qualname__rrr3rtupleboolrr6rDr^rr> __classcell__r<r<r:r=r"s<  O Kaf Implements Adam 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)} \\ &\hspace{13mm} \lambda \text{ (weight decay)}, \: \textit{amsgrad}, \:\textit{maximize}, \: \epsilon \text{ (epsilon)} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0\leftarrow 0 \text{ (second moment)},\: 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}\textbf{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}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}\textbf{if} \: amsgrad \\ &\hspace{10mm} v_t^{max} \leftarrow \mathrm{max}(v_{t-1}^{max},v_t) \\ &\hspace{10mm}\widehat{v_t} \leftarrow v_t^{max}/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\theta_t \leftarrow \theta_{t-1} - \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 `Adam: A Method for Stochastic Optimization`_. z Args: a 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 (L2 penalty) (default: 0) decoupled_weight_decay (bool, optional): if True, this optimizer is equivalent to AdamW and the algorithm will not accumulate weight decay in the momentum nor variance. (default: False) 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 a= .. Note:: A prototype implementation of Adam and AdamW for MPS supports `torch.float32` and `torch.float16`. .. _Adam\: A Method for Stochastic Optimization: https://arxiv.org/abs/1412.6980 .. _On the Convergence of Adam and Beyond: https://openreview.net/forum?id=ryQu7f-RZ )r&rXrYrZr[r\r_r`r+r]rarbr'r*r)r!r"r#r%c 'Csx|dur|dusJtjrJt| ts,Jt| ts:Jt| tsRJnt| } t| trn| j| jf| i}nd}t |D]\}}|s||n|| }||}||}||}tj s|rt }|jj |jj kr|jj |vsJd|d|d7}| dkrh|r|d| | nJ|rZt| trZ| jrJ||| }n|j|| d}n|j|| d}t|rt|}t|}t|}|rt||||<t|}|j}|dur|j}||f}||vr| j||dd||<||}n| }||d||r^t| tr^| jrB|jt|d| dn|| j||d| d n|| j||d| d |s|r|}|rt| tr| jrd| |} n d| |} n d| |} |rt| tr| jrd| |}!n d| |}!n d| |}!| | }"|"}#|!}$|r|rF||}%n||}%||t|%||||$|#||#}&n||$|#||#}&|r|||&n |||&nt|}d| |} d| |}!| | }"|!d }$|r*tj|||||d |||$|}&n||$|}&|j||&|" d |rzt||rzt||||<qzdS) NIIf capturable=True, params and state_steps must be on supported devices: .rralphaT)rBrA non_blocking)Zweight)value?)out) rJjit is_scriptingr0r3rrrBrA enumeratecompiler is_compilingrtypeZmul_rVZaddcmul_cloneaddrSZ view_as_realtoZlerp_ZsquarenegsqrtZcopy_maximumZadd_Zaddcdiv_rZview_as_complex)'r&rXrYrZr[r\r_r`r+r]rarbr'r*r)r!r"r#r% beta1_dictiparamrRrOrPZstep_tcapturable_supported_devicesrBrAkey device_beta1r>bias_correction1bias_correction2 step_sizeZ step_size_negbias_correction2_sqrtrQZdenomr<r<r=_single_tensor_adamYs                    rc +st|dkrdSttr:|s&tddkr:tdttrd|sPtddkrdtdttr|sztddkrtdtjs|rt d d t fd d t ||DsJd d|dur|dusJ|rJdt t ||||||g}ttr6tjdkr6jind}|D]P\\}}}}}}}ttt|}ttt|}ttt|}ttt|}ttt|} |dj}!|dur|!|vrΈj|!dd||!<|r||!n}"| r|r ttt|}#t|||||#nt|||||r*t|}tjs^| djr^tj| tjdddddn t| d| dkr|rt|d| n(|rtj||| dntj||| d}t||d|"t|ttjrt|d}$d}%n |}$d}%t||$||%~~$|rt| }&t| }'t |&dt |'dt!|'t"|&t#|&t$|'|&}(|'})|rttt|}#t%|#|t&|#}*n t&|}*t"|*|)t|*|t"|*|(t'|||*nfdd| D}&fdd| D}'t(fdd|&D}(dd|'D})|r`ttt|}#t%|#|t&|#}*n t&|}*t"|*|)t|*|t'|||*|(qBdS)Nrr,rr-zHbeta1 as a Tensor is not supported for capturable=False and foreach=TruezTensor beta1 must be 1-elementzHbeta2 as a Tensor is not supported for capturable=False and foreach=TruezTensor beta2 must be 1-elementF)Z supports_xlac3s.|]&\}}|jj|jjko$|jjvVqdSN)rBrx).0rNr>)rr<r= Tsz%_multi_tensor_adam..rkrlz#_foreach ops don't support autogradcpuTrBror/)rBrmcsg|]}dt|qSrrrr>)rar<r= sz&_multi_tensor_adam..csg|]}dt|qSrrr)rbr<r=rscsg|]}|dqS)r<rZbc)r'r<r=rcSsg|] }|dqS)rqr<rr<r<r=rr))rIr0rr7r2r1rJrvrwrallziprr"_group_tensors_by_device_and_dtypestrrBvaluesrlistr{rZ _foreach_negZis_cpu _foreach_add_rLZ _foreach_mul_Z _foreach_addZ_foreach_lerp_Z _foreach_mulZ_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&rXrYrZr[r\r_r`r+r]rarbr'r*r)r!r"r#r%grouped_tensorsrdevice_params_ device_grads_device_exp_avgs_device_exp_avg_sqs_Zdevice_max_exp_avg_sqs_device_state_steps__ device_params device_gradsdevice_exp_avgsdevice_exp_avg_sqsdevice_state_stepsrBrdevice_max_exp_avg_sqsZscaled_device_gradsrprrrrZexp_avg_sq_sqrtr<)rarbrr'r=_multi_tensor_adams                                        r)r&rXrYrZr[r\r_r`r+r]rarbr'r*r)r!r"r#r%returnc 'Cs|sdS|rtd|dur&|j|ini}|dur<|j|ini}t| trbt| jdkrb| j| ind}t||||||g}|D]:\\}}\\}}}}}}}tt t|}tt t|} tt t|}!tt t|}"tt t|}#d\}$}%|dur| ||j |dd}$|dur0| ||j |dd}%|dur^||vr^| j |dd||<||} t |#d|svt jnt j}&|&|| |!|"||#|| | | | |||$|%d|%durt |#|%gt|#qdS) Nz9Adam with fused=True does not support differentiable=Truer)NNT)rorr) r+r'rarbr*r)r!r_r`)r7rBr0rrrritemsrrrFr{rJrZ _fused_adam_Z _fused_adamw_rrI)'r&rXrYrZr[r\r_r`r+r]rarbr'r*r)r!r"r#r%Zgrad_scale_dictZfound_inf_dictZlr_dictrrBrrrrrrrrrrrrZdevice_grad_scaleZdevice_found_inffuncr<r<r= _fused_adams~$    r)Zsingle_tensor_fnF)r&rXrYrZr[r\r r"r#r$r_r`r]r%r+rarbr'r*r)r!cCs| dur8|dur8t||dd\}}|r8t|tr8|s8d}| durDd} |durPd}tjsttdd|Dsttd|rtj rtd| rtj rtd| rtj st }n|rtj st }nt }||||||||| ||||||||| | | d dS) znFunctional API that performs Adam algorithm computation. See :class:`~torch.optim.Adam` for details. NF)Z use_fusedcss|]}t|tjVqdSr)r0rJr)rtr<r<r=rszadam..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsz6torch.jit.script not supported with foreach optimizersz4torch.jit.script not supported with fused optimizers) r+r]rarbr'r*r)r!r"r#r_r`r%) rr0rrJrvrwrr7rsrtrrr)r&rXrYrZr[r\r r"r#r$r_r`r]r%r+rarbr'r*r)r!rrr<r<r=rqs\# )NFFNNNFF)%typingrrrrJrZ optimizerrrr r r r r rrrrrrrrrrrrr__all__r__doc__rrir3rrrrr<r<r<r=s Xr%H    G    u  `