Vula ikhodi ye-animation synthesis usebenzisa amanethiwekhi e-neural

Iqembu labacwaningi abavela eShanghai Technical University eshicilelwe amathuluzi Umlingisi, okuvumela ukusebenzisa izindlela zokufunda ngomshini ukulingisa ukunyakaza kwabantu kusetshenziswa izithombe ezimile, kanye nokushintsha izingubo, ukuzidlulisela kwenye indawo futhi kushintshe i-engeli lapho into ibonakala khona. Ikhodi ibhalwe ngePython
ngokusebenzisa uhlaka I-PyTorch. Umhlangano nawo udinga i-torchvision kanye ne-CUDA Toolkit.

Ikhithi yamathuluzi ithola isithombe esinezinhlangothi ezimbili njengokufaka futhi ihlanganise umphumela oshintshiwe ngokusekelwe kumodeli ekhethiwe. Izinketho ezintathu zokuguqulwa ziyasekelwa:
Ukudala into ehambayo elandela ukunyakaza lapho imodeli yaqeqeshwa khona. Ukudlulisa izici zokubukeka kusuka kumodeli kuya entweni (isibonelo, ukuguqulwa kwezingubo). Isizukulwane se-engeli entsha (isibonelo, ukuhlanganiswa kwesithombe sephrofayela ngokusekelwe esithombeni esigcwele ubuso). Zontathu lezi zindlela zingahlanganiswa, isibonelo, ungakha ividiyo kusuka esithombeni esilingisa ukusebenza kweqhinga eliyinkimbinkimbi le-acrobatic engutsheni ehlukene.

Ngesikhathi senqubo yokuhlanganisa, imisebenzi yokukhetha into esithombeni nokwenza izakhi zangemuva ezingekho lapho ihamba yenziwa kanyekanye. Imodeli yenethiwekhi ye-neural ingaqeqeshwa kanye futhi isetshenziselwe ukuguqulwa okuhlukahlukene. Okokulayisha iyatholakala amamodeli enziwe ngomumo akuvumela ukuthi usebenzise amathuluzi ngokushesha ngaphandle kokuqeqeshwa kokuqala. I-GPU enosayizi wenkumbulo okungenani ongu-8GB iyadingeka ukuze isebenze.

Ngokungafani nezindlela zokuguqula ezisuselwe ekuguqulweni ngamaphoyinti abalulekile achaza indawo yomzimba endaweni enezinhlangothi ezimbili, I-Lingisi izama ukuhlanganisa inetha elinezinhlangothi ezintathu ngencazelo yomzimba lisebenzisa izindlela zokufunda zomshini.
Indlela ehlongozwayo ivumela ukuguqulwa kucatshangelwa isimo somzimba womuntu siqu kanye nokuma kwamanje, kulingisa ukunyakaza kwemvelo kwezitho.

Ukugcina ulwazi lwangempela olunjengokwakheka, isitayela, imibala nokubonwa kobuso phakathi nenqubo yokuguqulwa, inethiwekhi ye-neural ye-adversarial ekhiqizayo (I-Liquid Warping GAN). Ulwazi mayelana nento engumthombo namapharamitha wokuhlonza kwayo okunembile kukhishwa ngokufaka isicelo inethiwekhi ye-convolutional neural.

Source: opennet.ru