o i@sDddlmZmZddlmZdddZddZd d Zdd d Zd S))Model normal_init)registry皙?c Cs.tdtt||||dddddd|id}|S)Nprecomputable_affine)nOnInFnP)Wbpad dropout_rate)initdimsparamsattrs)rforwardr)rr r r dropoutmodelrR/home/ubuntu/.local/lib/python3.10/site-packages/spacy/ml/_precomputable_affine.pyPrecomputableAffines  rcsdddddjjjdddd }jjfd |ddd ||jdf}jjd d|d<fd d}||fS)Nr rr r r rF)zerosT)trans2outrcs8|\}}|jdks J|jdksJ|j|jdks#J|jdt|||}||jdf}d|jdd||jdf}d}|f}j||jdf|}jj||d d }|f}|d }d |||jdfS) Nrrrrr )axis)rrrrTtrans1)rrrrr ) ndimshapeinc_grad&_backprop_precomputable_affine_paddingreshapesum transposeopsgemm)dY_idsdYidsXfWopfidXfdWopfir Xrr r rr rrbackward"s" "  zforward..backward) get_dim get_paramr)alloc2fr#r*r&xpsqueeze)rr3is_trainYfr4rr2rrs     $.%rc Csl|jd}|d}|d}|d}|jj|dkdd}|jj|||||dd}|d |||fS) Nrr r rfdtypeTr r)r#r5r)asarrayr*r&) rr,r-nBr r rmaskd_padrrrr%Js    r%Ncs drdrdSdddd}j|}jjd}j}t||jt |j d|d }t||jdd } d| d  d ||j d fd d}||j jdd|j7}|j|dd}|j d |fd d}||j jjdd|jd|j7}fdd} d} d} d} d}d t| D]A} | ||}jj |}jj |}t|d| kr|jj |} d|qt|| kr|8 d qdSdS)aThis is like the 'layer sequential unit variance', but instead of taking the actual inputs, we randomly generate whitened data. Why's this all so complicated? We have a huge number of inputs, and the maxout unit makes guessing the dynamics tricky. Instead we set the maxout weights to values that empirically result in whitened outputs given whitened inputs. r Nr rr r rg?)meanr rir<r=riig)locscalesizecs|dd}jj|jdfdd}||jdf}j|||||jdf}|7}j|}dkrQj|dS||dkS)Nrr<r=r) predictr)allocr#r& scatter_addflattenr?maxout)r-tokvecshiddensvectorsr rr rr rrrIs  zinit..predictg{Gz? ) has_paramr6anyr5r)alloc4fr7rr#floatr8sqrt set_paramrJrandomuniformr?normalrGr&copyrangevarrCabs)rr3Yr r rr)r-rNrItol_vartol_meant_maxt_iacts1r^rCrrQrrasP     $     r)r)NN) thinc.apirrutilrrrr%rrrrrs    8