o iA @sddlmZmZddlZddlmZddlmZmZmZm Z m Z m Z m Z m Z mZmZmZmZmZmZddgZGdddeZd d ed e d ed e d ed e_deedeedeedeedeedededededededededdfddZdeedeedeedeedeedededededededededdfddZe ed  !  ! !d$deedeedeedeedeeded"edBdedededededededdfd#dZdS)%)AnycastN)Tensor)_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported _foreach_doc!_get_capturable_supported_devices_get_scalar_dtype _maximize_doc _params_doc _to_scalar_use_grad_for_differentiable _view_as_real OptimizerParamsTAdadeltaadadeltacseZdZ     d ddddded eeBd ed ed ed edBdedededdffddZfddZde e e fde ede ede ede ede ef ddZ ed!ddZZS)"r??ư>rNF) capturablemaximizedifferentiableparamslrrhoeps weight_decayforeachrrrreturnc st|tr|dkrtdd|kstd|d|kr$dks,ntd|d|ks7td|d|ksBtd||||||||| d } t|| dS) NrzTensor lr must be 1-elementgzInvalid learning rate: rzInvalid rho value: zInvalid epsilon value: zInvalid weight_decay value: )rrrr rrr!r) isinstancernumel ValueErrorsuper__init__) selfrrrrr r!rrrdefaults __class__V/sda-disk/www/egybert/egybert_env/lib/python3.10/site-packages/torch/optim/adadelta.pyr's(  zAdadelta.__init__cst||jD]S}|dd|dd|dd|dd|dD]4}|j|g}t|dkr[t|ds[t |d}|drQtj |t |j d ntj |t d |d<q'q dS) Nr!rFrrrrstepdtypedevicer0) r& __setstate__ param_groups setdefaultstategetlentorch is_tensorfloattensorr r1)r(r6grouppp_statestep_valr*r,r-r3As&         zAdadelta.__setstate__r=params_with_gradgrads square_avgs acc_deltas state_stepsc Csd}|dD]n}|jdurq|t|O}|||jjr"td||j|j|} t| dkr_|drAtjdt |j dntjdt d| d <tj |tj d | d <tj |tj d | d <|| d || d || d q|S) NFrz*Adadelta does not support sparse gradientsrrr,r/r2r.) memory_format square_avg acc_delta) gradr9 is_complexappend is_sparse RuntimeErrorr6r8zerosr r1 zeros_likepreserve_format) r(r=rArBrCrDrE has_complexr>r6r,r,r- _init_groupTs2        zAdadelta._init_groupcCs|d}|dur!t |}Wdn1swY|jD]J}g}g}g}g}g}|d|d|d|d|d|d|d|d f\} } } } } }}}|||||||}t|||||| | | | | ||||d q$|S) zPerform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nrrrr r!rrr) rrrr r!rrrrQ) _cuda_graph_capture_health_checkr9 enable_gradr4rRr)r(closurelossr=rArBrCrDrErrrr r!rrrrQr,r,r-r.sd    z Adadelta.step)rrrrNN)__name__ __module__ __qualname__rr;rboolr'r3dictstrrlistrRrr. __classcell__r,r,r*r-rs^      $  +aImplements Adadelta algorithm. .. math:: \begin{aligned} &\rule{110mm}{0.4pt} \\ &\textbf{input} : \gamma \text{ (lr)}, \: \theta_0 \text{ (params)}, \: f(\theta) \text{ (objective)}, \: \rho \text{ (decay)}, \: \lambda \text{ (weight decay)} \\ &\textbf{initialize} : v_0 \leftarrow 0 \: \text{ (square avg)}, \: u_0 \leftarrow 0 \: \text{ (accumulate variables)} \\[-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} v_t \leftarrow v_{t-1} \rho + g^2_t (1 - \rho) \\ &\hspace{5mm}\Delta x_t \leftarrow \frac{\sqrt{u_{t-1} + \epsilon }}{ \sqrt{v_t + \epsilon} }g_t \hspace{21mm} \\ &\hspace{5mm} u_t \leftarrow u_{t-1} \rho + \Delta x^2_t (1 - \rho) \\ &\hspace{5mm}\theta_t \leftarrow \theta_{t-1} - \gamma \Delta x_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 `ADADELTA: An Adaptive Learning Rate Method`_. z Args: ar lr (float, Tensor, optional): coefficient that scale delta before it is applied to the parameters (default: 1.0) rho (float, optional): coefficient used for computing a running average of squared gradients (default: 0.9). A higher value of `rho` will result in a slower average, which can be helpful for preventing oscillations in the learning process. eps (float, optional): term added to the denominator to improve numerical stability (default: 1e-6). weight_decay (float, optional): weight decay (L2 penalty) (default: 0) z zd .. _ADADELTA\: An Adaptive Learning Rate Method: https://arxiv.org/abs/1212.5701 rrBrCrDrErrrr rrrrQr"csftjs$| r$tddtfddt||ddDs$tddtjs-t |}t|||||ddD]y\} }}}}|d 7}| sF|n| }|d krT|j | |d }t | rht |}t |}t |}| |j||d |d | |}| |}| r|}|| || |j||d |d t | rt|}| j|| d q7dS) NF supports_xlac30|]\}}|jj|jjko|jjvVqdSrWr1type.0r>r.capturable_supported_devicesr,r-    z*_single_tensor_adadelta..TstrictIIf capturable=True, params and state_steps must be on supported devices: .rralphavalue)r9compiler is_compilingr allzipAssertionErrorjit is_scriptingraddrJ view_as_realmul_addcmul_sqrt_clonediv_view_as_complexadd_)rrBrCrDrErrrr rrrrQparamrIrGrHr.stddeltar,rgr-_single_tensor_adadeltasD           rcs:| rtdtjs*| r*tddtfddt||ddDs*tdd t|d kr2dSt|}t |||||g} | D]\\}}}}}}t t t|}t t t|}t t t|}t t t|}t t t|}| ryt||||tjs|d jrtj|tjd d d d dnt|d| rt|}|d kr| rtj|||dntj|||d}t||tj|||d|dt||}t|t||}t|t||t||t||tj|||d|d| rt|tjrt|| t||qDtj||| dqDdS)Nz#_foreach ops don't support autogradFr`c3rbrWrcrergr,r-riIrjz)_multi_tensor_adadelta..Trkrmrnrrcpu)r1rorrq)rwr9rsrtr rurvr8rr"_group_tensors_by_device_and_dtypevaluesrr^rris_cpu _foreach_add_r< _foreach_neg _foreach_add _foreach_mul__foreach_addcmul__foreach_sqrt_ _foreach_div_r#)rrBrCrDrErrrr rrrrQgrouped_tensorsdevice_params_ device_grads_device_square_avgs_device_acc_deltas_device_state_steps__ device_params device_gradsdevice_square_avgsdevice_acc_deltasdevice_state_stepsrdeltasr,rgr-_multi_tensor_adadelta1s                 r)single_tensor_fnFr!c Cstjstdd|Dstd|durt||dd\}}|r*tjr*td|r4tjs4t}nt }||||||| | | | | |||d dS) zvFunctional API that performs Adadelta algorithm computation. See :class:`~torch.optim.Adadelta` for details. css|] }t|tjVqdSrW)r#r9r)rftr,r,r-ris  zadadelta..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF) use_fusedz6torch.jit.script not supported with foreach optimizers)rrrr rrrrQ) r9rsrtrurMrrxryrr)rrBrCrDrErr!rrQrrrr rrfuncr,r,r-rs<  )FNFF)typingrrr9r optimizerrrrr r r r r rrrrrr__all__r__doc__r^r;r[rrrr,r,r,r-s @&  5      <      j