a d8@s:ddlZddlZddlmZddlmZmZmZmZmZm Z m Z m Z ddl m Z mZddlmZddgZGd ddeZd d je e d e_de ee ee ee ee eeeeeeeeed ddZe ee ee ee ee eeeeeeed ddZe ee ee ee ee eeeeeeed ddZdS)N)Tensor) Optimizer_use_grad_for_differentiable _get_value_dispatch_sqrt_stack_if_compiling_default_to_fused_or_foreach_differentiable_doc _foreach_doc)ListOptional)"_group_tensors_by_device_and_dtypeRAdamradamcsTeZdZddddeeedfdd Zfd d Zd d ZedddZ Z S)rMbP?g?g+?:0yE>rNFforeachdifferentiablec 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: {})lrbetaseps weight_decayrr) ValueErrorformatdictsuper__init__) selfparamsrrrrrrdefaults __class__Z/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/torch/optim/radam.pyr!s& zRAdam.__init__cst||jD]}|dd|ddqt|j}t|dkoZt |dd}|s|D]}t t |d|d<qddS)NrrFrstep) r __setstate__ param_groups setdefaultliststatevalueslentorchZ is_tensortensorfloat)r"r.groupZ state_valuesZstep_is_tensorsr%r'r(r*-s    zRAdam.__setstate__c Cs|dD]}|jdur|||jjr0td||j|j|}t|dkrtd|d<tj|tj d|d<tj|tj d|d<||d||d||dqdS) Nr#z'RAdam does not support sparse gradientsrrr)Z memory_formatexp_avg exp_avg_sq) gradappendZ is_sparse RuntimeErrorr.r0r1r2Z zeros_likepreserve_format) r"r4params_with_gradgradsexp_avgs exp_avg_sqs state_stepspr.r'r'r( _init_group:s$        zRAdam._init_groupc Csd}|dur:t|}Wdn1s00Y|jD]l}g}g}g}g}g}|d\} } |||||||t|||||| | |d|d|d|d|dd q@|S) zPerforms a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. Nrrrrrr)beta1beta2rrrrr)r1Z enable_gradr+rCr) r"closureZlossr4r=r>r?r@rArDrEr'r'r(r)Ss6 $  z RAdam.step)rrrr)N) __name__ __module__ __qualname__r boolr!r*rCrr) __classcell__r'r'r%r(r s a Implements RAdam 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)}, \: \lambda \text{ (weightdecay)}, \\ &\hspace{13mm} \epsilon \text{ (epsilon)} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0 \leftarrow 0 \text{ ( second moment)}, \\ &\hspace{18mm} \rho_{\infty} \leftarrow 2/(1-\beta_2) -1 \\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{6mm}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{6mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{6mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{6mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\ &\hspace{6mm}\rho_t \leftarrow \rho_{\infty} - 2 t \beta^t_2 /\big(1-\beta_2^t \big) \\[0.1.ex] &\hspace{6mm}\textbf{if} \: \rho_t > 5 \\ &\hspace{12mm} l_t \leftarrow \frac{\sqrt{ (1-\beta^t_2) }}{ \sqrt{v_t} +\epsilon } \\ &\hspace{12mm} r_t \leftarrow \sqrt{\frac{(\rho_t-4)(\rho_t-2)\rho_{\infty}}{(\rho_{\infty}-4)(\rho_{\infty}-2) \rho_t}} \\ &\hspace{12mm}\theta_t \leftarrow \theta_{t-1} - \gamma \widehat{m_t} r_t l_t \\ &\hspace{6mm}\textbf{else} \\ &\hspace{12mm}\theta_t \leftarrow \theta_{t-1} - \gamma \widehat{m_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 `On the variance of the adaptive learning rate and beyond`_. This implementation uses the same weight_decay implementation as Adam (were the weight_decay is applied to the gradient) and not the one from AdamW (were weight_decay is applied to the update). This is different from the `author's implementation`_. a Args: params (iterable): iterable of parameters to optimize or dicts defining parameter groups lr (float, optional): learning rate (default: 1e-3) 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) {foreach} {differentiable} .. _On the variance of the adaptive learning rate and beyond: https://arxiv.org/abs/1908.03265 .. _author's implementation: https://github.com/LiyuanLucasLiu/RAdam rF) r#r>r?r@rArrrDrErrrc Cstdd|Dstd|dur4t||dd\} }|rJtjrJtd|r^tjs^t} nt} | |||||||| | | |d dS) zpFunctional API that performs RAdam algorithm computation. See :class:`~torch.optim.RAdam` for details. css|]}t|tjVqdS)N) isinstancer1r).0tr'r'r( zradam..zPAPI has changed, `state_steps` argument must contain a list of singleton tensorsNF)Z use_fusedz6torch.jit.script not supported with foreach optimizers)rDrErrrr)allr;r r1ZjitZ is_scripting_multi_tensor_radam_single_tensor_radam)r#r>r?r@rArrrDrErrr_funcr'r'r(rs0) r#r>r?r@rArDrErrrrcCspt|D]`\} } || } || }|| }|| }|d7}t|}d||}d||}|dkrp| j| |d} ||j| d|d||j| | d|d||}dd|d}|d||||}|dkrXt|d|d||d|d|}|}| r$|| }n || }t||}| j||||ddq| j||ddqdS) Nrralpha)valueg@g) enumerateraddZmul_Zadd_Zaddcmul_mathsqrt)r#r>r?r@rArDrErrrriparamr9r7r8Zstep_tr)bias_correction1bias_correction2Zbias_corrected_exp_avgrho_infrho_trectexp_avg_sq_sqrtZ adaptive_lrr'r'r(rSsB     rScst|dkrdS| rJdt|||||g} | D]p\} } }}}t|ddddfdd|D}fdd|D}fdd|D}|dkrtj| | |d } t|tj|| dd t|t|| | dfd d|D}d d|D}t|}t|| d d|D}t ||}t fd dt ||D}t | |||dd|D}t fddt ||D}t | |||q6dS)Nrz#_foreach ops don't support autogradrrYcs8g|]0}dt|t|dt|qS)rYrrrMr))rErcr'r( Fsz'_multi_tensor_radam..csg|]}dt|qSrrgrh)rDr'r(riIrPcsg|]}dt|qSrjrgrh)rEr'r(riJrPrVcsDg|]<}|dkrr?r@rArDrErrrrZgrouped_tensorsZgrouped_paramsZ grouped_gradsZgrouped_exp_avgsZgrouped_exp_avg_sqsZgrouped_state_stepsZ rho_t_listrarbreZ unrectifiedrfZbias_correction_sqrtZdenomZ step_sizer')rDrErrcr(rR*s>           rR)NF)r]r1rZ optimizerrrrrrr r r typingr r Ztorch.utils._foreach_utilsr__all__rr__doc__rJr3rrSrRr'r'r'r(sh ( o'D 6 >