Kwanan nan NVIDIA ta saki lambar tushe don StyleGAN3, tsarin koyon injin da ya dogara da hanyoyin sadarwar jijiyoyin jijiyoyin jiki (GAN) don haɗa hotunan zahiri na fuskokin mutane.
A cikin StyleGAN3 suna samuwa don saukar da shirye-shiryen horar da samfuran da aka horar a cikin tarin Flickr-Faces-HQ (FFHQ), wanda ya ƙunshi hotunan PNG dubu 70 na manyan fuskokin mutane (1024 × 1024). Bugu da kari, akwai samfuran da aka gina akan AFHQv2 (hotunan fuskokin dabbobi) da Fakitoci (hotunan fuskokin mutane daga hotunan zane na gargajiya).
Game da StyleGAN3
Zane yana mai da hankali kan fuskoki, amma ana iya horar da tsarin don samar da kowane irin abu, kamar shimfidar wurare da motoci. Menene ƙari, ana ba da kayan aikin don koyan kai na cibiyar sadarwa na jijiyoyi ta amfani da tarin hotunan ku. Yana buƙatar katunan zane na NVIDIA ɗaya ko fiye (An ba da shawarar Tesla V100 ko A100 GPUs), aƙalla 12GB na RAM, PyTorch 1.9, da CUDA 11.1+ Kayan aiki. Don tantance yanayin wucin gadi na fuskokin da aka karɓa, ana haɓaka na'urar ganowa ta musamman.
Tsarin yana ba da damar haɗa hoto na sabon fuska dangane da haɗaɗɗun fasallan fuskoki da yawa, haɗa halayen su na asali, ban da daidaita hoto na ƙarshe zuwa shekarun da ake buƙata, jinsi, tsawon gashi, halayyar murmushi, siffar hanci, launin fata, tabarau, kusurwar hoto.
Generator yana ɗaukar hoton azaman tarin salo, yana rarrabe cikakkun bayanan halayen ta atomatik (ƙulle-ƙulle, gashi, tabarau) na manyan halayen gabaɗaya (matsayi, jinsi, canje-canjen da suka shafi shekaru) kuma yana ba su damar haɗa kai ba tare da izini ba tare da ma'anar manyan kadarori ta hanyar abubuwan nauyi kuma a sakamakon haka, ana samar da hotuna waɗanda a bayyane ba za a iya bambanta su da ainihin hotunan ba.
Sigar farko na fasahar StyleGAN (wanda aka saki a cikin 2019), sannan ingantaccen sigar StyleGAN2 a 2020, wanda ke inganta ingancin hoto da cire wasu kayan tarihi. A lokaci guda, tsarin ya kasance a tsaye, wato bai yarda da raye -raye na zahiri ko motsi na fuska ba. Lokacin haɓaka StyleGAN3, babban burin shine daidaita fasahar don amfani a cikin raye -raye da bidiyo.
StyleGAN3 yana amfani da tsarin zane-zanen hoto da ba a sake baay yana ba da sabbin yanayin horo na cibiyar sadarwa kuma ya haɗa da sabbin abubuwan amfani don hangen nesa (visualizer.py), bincike (avg_spectra.py) da tsara bidiyo (gen_video.py). Har ila yau aiwatarwa yana rage yawan amfani da ƙwaƙwalwa kuma yana hanzarta tsarin koyo.
Babban fasali na gine -gine na StyleGAN3 shine sauyawa zuwa fassarar duk sigina a cikin cibiyar sadarwa na jijiyoyin jiki ta hanyar ci gaba da aiwatarwa, wanda ya ba da damar yin amfani da matsayin dangi ta hanyar samar da sassa, ba a daura da cikakken daidaiton kowane pixels a cikin hoto, amma an gyara shi a saman abubuwan da aka wakilta.
Duk da yake a cikin StyleGAN da StyleGAN2, ɗaukar hoto zuwa pixels yayin gini ya haifar da matsaloli tare da ma'ana mai ƙarfiMisali, lokacin da hoton ke motsawa, akwai rashin daidaituwa na ƙananan bayanai, kamar wrinkles da gashi, waɗanda da alama suna motsi daban da sauran hoton fuskar, ban da wannan a cikin StyleGAN3 an warware waɗannan matsalolin kuma fasaha tana da zama sosai dace ga video tsara.
A ƙarshe, Har ila yau, daraja ambata sanarwar kirkirar NVIDIA da Microsoft na mafi girman samfurin yare na MT-NLG dangane da hanyar sadarwa mai zurfi mai zurfi tare da »gine -gine mai canzawa.
Samfurin ya ƙunshi sigogin biliyan 530 kuma an yi amfani da tafkin 4480 GPUs don horo (sabobin 560 DGX A100 tare da 8 A100 GPUs na 80 GB kowannensu). Yankunan aikace -aikacen ƙirar ana kiran su warware matsalar sarrafa bayanai a cikin yaren halitta, kamar yin annabcin kammala jumla da ba a ƙare ba, amsa tambayoyi, fahimtar karatu, samar da ƙarshe a cikin yaren halitta, da yin nazarin shubuhar ma'anar kalmomi..
Idan kuna sha'awar ƙarin sani game da shi, zaku iya duba cikakkun bayanai na StyleGAN3 A cikin mahaɗin mai zuwa.