U Jc@slUddlmZmZmZmZddlZddlmmZ ddlm Z gZ ee e d<ejjGdddeZdS))ListDictOptionalTupleN)Tensor__all__c @sdeZdZdeeeeeefeeeeeeed ddZee ed d d Z ee ed d dZ dS)_FunctionalAdamMbP?g?g+?:0yE>F) paramslrbetaseps weight_decayamsgradmaximizeforeachfused_allow_empty_param_listc Csd|kstd|d|ks,td|d|dkrDdksXntd|dd|dkrpdksntd|dd|kstd ||||d|d|d |_||_||_||_| |_tj t tj t t tj ffi|_ t|dkr | s td d |i|_dS) Nr zInvalid learning rate: {}zInvalid epsilon value: {}rg?z%Invalid beta parameter at index 0: {}z%Invalid beta parameter at index 1: {}zInvalid weight_decay value: {})rrbeta1beta2rz%optimizer got an empty parameter listr ) ValueErrorformatdefaultsrrrrtorchjitZannotaterrstrstatelen param_group) selfr rrrrrrrrrr$K/tmp/pip-unpacked-wheel-gikjz4vx/torch/distributed/optim/functional_adam.py__init__s0 $z_FunctionalAdam.__init__)paramgradc CsZ|g}g}g}g}g}g}g} |dk r:||||||jkri|j|<|j|} td| d<tj|tjd| d<tj|tjd| d<|jrtj|tjd| d<|j|} || d|| d|jr|| d| | dtXtj |||||| |j|j |j d|j d |j d |j d |j d |j |j ddd W5QRXdS)zo Similar to step, but operates on a single parameter and optionally a gradient tensor. Nr stepZ memory_formatexp_avg exp_avg_sqmax_exp_avg_sqrrrrr rrrrrrrrrZ grad_scaleZ found_inf)appendr rtensor zeros_likepreserve_formatrno_gradFadamrrrr) r#r'r(r params_with_gradgradsexp_avgs exp_avg_sqsmax_exp_avg_sqs state_stepsr r$r$r% step_param@sV       z_FunctionalAdam.step_param) gradientsc Cs|jd}g}g}g}g}g}g}t|t|krXtddt|ddt|t|jd|D]\} } | dk rh|| || | |jkri|j| <|j| } td| d<tj| tj d| d <tj| tj d| d <|j rtj| tj d| d <|j| } || d || d |j r6|| d || dqht Xt j |||||||j |j|jd |jd |jd|jd|jd|j|jdddW5QRXdS)Nr zEthe gradients passed in does not equal to the size of the parameters!zParams length: z. zGradients length: r r)r*r+r,r-rrrrrr.)r"r!rzipr/r rr0r1r2rr3r4r5rrrr) r#r=r r6r7r8r9r:r;r'Zgradientr r$r$r%r)ssh         z_FunctionalAdam.stepN) r r r r FFFFF) __name__ __module__ __qualname__rrfloatrboolr&rr<r)r$r$r$r%rs.  ,3r)typingrrrrrZtorch.optim._functionalZoptimZ _functionalr4rrr__annotations__rscriptobjectrr$r$r$r%s