U yh@sfUddlmZmZmZddlZddlmmZddlm Z gZ ee e d<ej jGdddZdS))DictListOptionalN)Tensor__all__c @s\eZdZdeeeeeeeeeeed ddZeeeddd Z eeed d d Z d S)_FunctionalSGD{Gz?F) paramslrmomentum dampening weight_decaynesterovmaximizeforeachfused_allow_empty_param_listc Csr||||d|_||_||_||_| |_tjttj tt tj ffi|_ t |dkrd| sdt dd|i|_dS)N)r r r rrz%optimizer got an empty parameter listr )defaultsrrrrtorchjitZannotaterrstrstatelen ValueError param_group) selfr r r r rrrrrrrX/var/www/html/venv/lib/python3.8/site-packages/torch/distributed/optim/functional_sgd.py__init__s$z_FunctionalSGD.__init__)paramgradc Cs|jd}|jd}|jd}|jd}|g}g}g} d} |dk r| ||jrVd} ||jkrji|j|<|j|} d| kr|dn|| dt4tj|| ||||||j|j | |j |j ddd W5QRX|j|} |d } | dk r| | d<dS) z[Similar to self.step, but operates on a single parameter and its gradient. rr r r FNTmomentum_buffer rr r r rrhas_sparse_gradrrZ grad_scaleZ found_infr) rappend is_sparserrno_gradFsgdrrrr) rr r!rr r r r momentum_buffer_listgradsr$rr"rrr step_param6sN           z_FunctionalSGD.step_param) gradientscCsp|jd}g}g}g}|jd}|jd}|jd}|jd} t|t|krttddt|dd t|d } t||D]n\} } | dk r|| || | jrd } | |jkri|j| <|j| } d | kr|dq|| d qt 4t j ||||||| |j |j | |j|jddd W5QRXt|D].\}}|j|} ||}|dk r<|| d <qs