o
    io@                  $   @   s  d dl mZ d dlZd dlmZ ddlmZmZmZmZm	Z	m
Z
mZmZmZmZmZmZmZmZmZ ddgZG dd deZd	e d
e	 de de de de_dee dee dee dee dee dee dedededededededededdf ddZdee dee dee dee dee dee dedededededededededdf dd Zeed!		"	"	"	"d%dee dee dee dee dee dee d#edB dedededededededededdf"d$dZdS )&    )castN)Tensor   )_capturable_doc_default_to_fused_or_foreach_differentiable_doc_disable_dynamo_if_unsupported_foreach_doc!_get_capturable_supported_devices_get_scalar_dtype
_get_value_maximize_doc_params_doc
_to_scalar_use_grad_for_differentiable_view_as_real	OptimizerParamsTASGDasgdc                       s   e Zd Z									dded	eeB d
edededededB dedededdf fddZ fddZdd Z	e
dddZ  ZS )r   {Gz?-C6?      ?    .Ar   NFparamslrlambdalphat0weight_decayforeachmaximizedifferentiable
capturablereturnc              
      st   t |tr| dkrtdd|kstd| d|ks%td| ||||||||	|
d	}t || d S )Nr   zTensor lr must be 1-elementg        zInvalid learning rate: zInvalid weight_decay value: )	r   r   r   r   r   r    r!   r"   r#   )
isinstancer   numel
ValueErrorsuper__init__)selfr   r   r   r   r   r   r    r!   r"   r#   defaults	__class__ R/sda-disk/www/egybert/egybert_env/lib/python3.10/site-packages/torch/optim/asgd.pyr)      s"   zASGD.__init__c                    s   t  | | jD ]q}|dd  |dd |dd |dd |d D ]R}| j|g }t|dkryt|d sOt	|d }tj
|t |jd	|d< t|d
 sdtj
|d
 t |jd	|d
< t|d sytj
|d t |jd	|d< q'q	d S )Nr    r!   Fr"   r#   r   r   step)dtypedeviceetamu)r(   __setstate__param_groups
setdefaultstategetlentorch	is_tensorfloattensorr   r2   )r*   r8   grouppp_statestep_valr,   r.   r/   r5   ?   s2   




zASGD.__setstate__c                 C   s  d}|d D ]~}	|	j d ur|t|	O }||	 |	j jr!td||	j  | j|	 }
t|
dkrhtjd|	j	t
 d|
d< tjt|d |	j	t
 d  |
d	< tjd|	j	t
 d|
d
< tj|	tjd|
d< ||
d
  ||
d  ||
d	  ||
d  q|S )NFr   z&ASGD does not support sparse gradientsr   r.   )r2   r1   r0   r   r3   r4   )memory_formatax)gradr;   
is_complexappend	is_sparseRuntimeErrorr8   r:   zerosr2   r   	as_tensorr   clonedetachones
zeros_likepreserve_format)r*   r?   params_with_gradgradsmusaxsetasstate_stepshas_complexr@   r8   r.   r.   r/   _init_groupW   sB   





	


zASGD._init_groupc                 C   s   |    d}|dur!t  | }W d   n1 sw   Y  | jD ]?}g }g }g }g }g }g }	| |||||||	}
t||||||	|d |d |d |d |d |d |d |d	 |d
 |
d q$|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   r   r   r   r   r    r!   r"   r#   rW   ) _cuda_graph_capture_health_checkr;   enable_gradr6   rX   r   )r*   closurelossr?   rQ   rR   rS   rT   rU   rV   rW   r.   r.   r/   r0   }   sF   

z	ASGD.step)	r   r   r   r   r   NFFFN)__name__
__module____qualname__r   r=   r   boolr)   r5   rX   r   r0   __classcell__r.   r.   r,   r/   r      sJ    	
!&zImplements Averaged Stochastic Gradient Descent.

    It has been proposed in `Acceleration of stochastic approximation by
    averaging`_.

    Args:
        am  
        lr (float, Tensor, optional): learning rate (default: 1e-2)
        lambd (float, optional): decay term (default: 1e-4)
        alpha (float, optional): power for eta update (default: 0.75)
        t0 (float, optional): point at which to start averaging (default: 1e6)
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
        z	
        z

    .. _Acceleration of stochastic approximation by averaging:
        https://meyn.ece.ufl.edu/wp-content/uploads/sites/77/archive/spm_files/Courses/ECE555-2011/555media/poljud92.pdf

    r   rR   rT   rS   rU   rV   r   r   r   r   r   r!   r"   r#   rW   r$   c       	      
   C   s"  t j s	t|}t| D ] \}}|| }|s|n| }|| }|| }|| }|| }t j s^|r^t }|jj	|jj	  krN|jj	  krN|jj	krVn n|jj	|v s^t
d| dt |rrt |}t |}t |}|d7 }|
dkr|j||
d}|r|d||   |j||dd nt|}|d||   |j|| d |s| dkr|||| n|| |r||d|| |  |	   |dt || t |  qt|}t |d|| |  |	  }|| t dtd||  }|| qd S )NUIf capturable=True, params, mus, etas, and state_steps must be on supported devices: .r   r   r   value)r;   jitis_scriptingr   	enumeratecompileris_compilingr
   r2   typeAssertionErrorrF   view_as_realaddmul_addcmul_r   add_itemsubcopy_maximum	ones_likerK   max)r   rR   rT   rS   rU   rV   r   r   r   r   r   r!   r"   r#   rW   iparamrE   r4   rD   r3   step_tcapturable_supported_devices	eta_valuer0   new_etanew_mur.   r.   r/   _single_tensor_asgd   sb   






"
r   c       	       	      s  t | dkrd S |rtdtj s4|r4tddtfddt| |||ddD s4td	 d
tt	
| |||||g}| D ]\\}\\}}}}}}}ttt |}ttt |}ttt |}ttt |}ttt |}ttt |}|rt||| |rt|}tj s|d jrtj|tjddddd nt|d |
dkr|rtj|||
d |}ntj|||
d}tj||d ntj||d}tj|||dd ~t||}t||| ~|r4t|}t|d t| t|| ~t|}t| t|d t|  t| t| t|| qG fdd|D }fdd|D }t|| t|| qGd S )Nr   z#_foreach ops don't support autogradF)supports_xlac                 3   sV    | ]&\}}}}|j j|j j  ko|j j  ko|j jkn  o&|j j v V  qd S r]   )r2   rn   ).0r@   r4   r3   r0   )r~   r.   r/   	<genexpr>3  s    
2

z%_multi_tensor_asgd.<locals>.<genexpr>T)strictrc   rd   g      ?cpur2   re   r   rf   rg   c                    s.   g | ]}t jd  |     dqS r   r   )r;   rK   r   r0   )r   r2   r   r   r.   r/   
<listcomp>  s     z&_multi_tensor_asgd.<locals>.<listcomp>c                    s,   g | ]}t jd td t|   dqS r   )r;   rK   rz   r   r   )r2   r   r.   r/   r     s    )r:   ro   r;   rl   rm   r
   allzipr   r   "_group_tensors_by_device_and_dtypeitemsr   listr   r   _foreach_negis_cpu_foreach_add_r>   _foreach_add_foreach_addcmul__foreach_sub_foreach_maximum__foreach_reciprocal__foreach_copy__foreach_mul_foreach_mul__foreach_pow_) r   rR   rT   rS   rU   rV   r   r   r   r   r   r!   r"   r#   rW   grouped_tensors_grouped_params_grouped_grads_grouped_axs_grouped_mus_grouped_etas_grouped_state_steps_grouped_paramsgrouped_gradsgrouped_axsgrouped_musgrouped_etasgrouped_state_stepsintermediatenew_musnew_etasr.   )r   r~   r2   r   r   r   r/   _multi_tensor_asgd  s   



r   )single_tensor_fnFr    c                C   sr   |du rt | |dd\}}|rtj rtd|r"tj s"t}nt}|| |||||||||||||	|
d dS )znFunctional API that performs asgd algorithm computation.

    See :class:`~torch.optim.ASGD` for details.
    NF)	use_fusedz6torch.jit.script not supported with foreach optimizers)	r   r   r   r   r   r!   r"   r#   rW   )r   r;   ri   rj   rI   r   r   )r   rR   rT   rS   rU   rV   r    r!   r"   r#   rW   r   r   r   r   r   r   funcr.   r.   r/   r     s4   

)NFFFF)typingr   r;   r   	optimizerr   r   r   r   r	   r
   r   r   r   r   r   r   r   r   r   __all__r   __doc__r   r=   ra   r   r   r   r.   r.   r.   r/   <module>   s   D 
	

Q	

 
	
