Vula ikhowudi ye-animation synthesis usebenzisa inethiwekhi ze-neural

Iqela labaphandi abavela kwiYunivesithi yaseShanghai yobuGcisa ipapashiwe izixhobo Umlingisi, okuvumela ukusebenzisa iindlela zokufunda umatshini ukulinganisa ukunyakaza kwabantu usebenzisa imifanekiso engatshintshiyo, kunye nokutshintsha iimpahla, ukuhambisa kwenye indawo kwaye utshintshe i-angle apho into ibonakala khona. Ikhowudi ibhalwe kwiPython
usebenzisa isakhelo I-PyTorch. INdibano nayo ifuna torchvision kunye ne-CUDA Toolkit.

Isixhobo se-toolkit sifumana umfanekiso omacala mabini njengegalelo kwaye idibanise iziphumo ezilungisiweyo ngokusekelwe kumzekelo okhethiweyo. Iinketho ezintathu zenguqu ziyaxhaswa:
Ukudala into ehambayo elandela iintshukumo apho imodeli yaqeqeshwa khona. Ukudlulisa izinto zembonakalo ukusuka kwimodeli ukuya kwinto (umzekelo, utshintsho lwempahla). Isizukulwana se-engile entsha (umzekelo, ukuhlanganiswa komfanekiso weprofayili ngokusekelwe kwifoto yobuso obugcweleyo). Zontathu iindlela zinokudityaniswa, umzekelo, unokuvelisa ividiyo kwifoto elingisa ukusebenza kweqhinga elintsonkothileyo le-acrobatic kwiimpahla ezahlukeneyo.

Ngexesha lenkqubo yokudibanisa, imisebenzi yokukhetha into kwifoto kunye nokwenza izinto ezingasemva ezingekhoyo xa ushukuma zenziwa ngaxeshanye. Imodeli yenethiwekhi ye-neural inokuqeqeshwa kube kanye kwaye isetyenziselwe iinguqu ezahlukeneyo. Yokulayisha iyafumaneka iimodeli esele zilungile ezikuvumela ukuba usebenzise ngokukhawuleza izixhobo ngaphandle koqeqesho lokuqala. I-GPU enobungakanani bememori ubuncinci obuyi-8GB iyafuneka ukuze isebenze.

Ngokungafaniyo neendlela zokuguqulwa ezisekelwe kwinguqu ngamanqaku aphambili achaza indawo yomzimba kwindawo ene-dimensional, i-Impersonator izama ukudibanisa i-mesh-dimensional mesh kunye nenkcazo yomzimba usebenzisa iindlela zokufunda umatshini.
Indlela ecetywayo ivumela ukuguqulwa kuthathelwe ingqalelo ubume bomzimba womntu kunye nokuma kwangoku, ukulinganisa ukunyakaza kwendalo kwamalungu.

Ukugcina ulwazi lwangaphambili olunje ngemibala, isitayile, imibala kunye nokuqondwa kobuso ngexesha lenkqubo yotshintsho, inethiwekhi ye-neural ye-adversarial (Ulwelo Warping GAN). Ulwazi malunga nomthombo wenjongo kunye neeparameters zokuchongwa kwayo okuchanekileyo zicatshulwa ngokufaka isicelo inethiwekhi ye-convolutional neural.

umthombo: opennet.ru