Mva nje I-NVIDIA ikhuphe ikhowudi yemvelaphi ye-StyleGAN3, inkqubo yokufunda ngomatshini esekwe kuthungelwano olubi lwe-neural network (GAN) ukwenza imifanekiso eyiyo yobuso babantu.
KwisitayileGAN3 ziyafumaneka ukukhuphela iimodeli ezilungele ukusetyenziswa eziqeqeshwe kwingqokelela IFlickr-Faces-HQ (FFHQ), ebandakanya imifanekiso engama-70 amawaka e-PNG yobuso obuphezulu bomntu (1024 × 1024). Ukongeza, kukho iimodeli ezakhelwe kwisiseko se-AFHQv2 (iifoto zobuso bezilwanyana) kunye neMetfaces (imifanekiso yobuso babantu abavela kwimifanekiso yepeyinti yakudala) ingqokelela.
Malunga nesimboGAN3
Ukuyila ijolise kubuso, kodwa inkqubo ingaqeqeshwa ukuvelisa naluphi na uhlobo lwento, njengemimandla kunye neemoto. Yintoni egqithisile, izixhobo zibonelelwa ngokuzifundela kwenethiwekhi ye-neural usebenzisa eyakho ingqokelela yemifanekiso. Ifuna ikhadi elinye lemizobo yeNVIDIA (I-Tesla V100 okanye i-A100 GPU iyacetyiswa), ubuncinci i-12GB ye-RAM, iPyTorch 1.9, kunye ne-CUDA 11.1+ Toolkit. Ukuchonga ubunjani bobuso obufunyenweyo, umatshini okhethekileyo uyenziwa.
Inkqubo ivumela ukuhlanganisa umfanekiso wobuso obutsha ngokusekwe kukudityaniswa kweempawu zobuso obuninziUkudibanisa amanqaku abo, ukongeza kulungelelwaniso lomfanekiso wokugqibela kubudala obufunekayo, isini, ubude beenwele, uncumo, ukumila kwempumlo, umbala wolusu, iiglasi, iangile yokufota.
Umvelisi uphatha umfanekiso njengokuqokelelwa kwezitayile, ngokwahlulahlula iinkcukacha zomntu (amabala, iinwele, iiglasi) zeempawu ezikumgangatho ophezulu ngokubanzi (ukuma komzimba, isini, utshintsho olunxulumene nobudala) kwaye ivumela ukuba zidityaniswe ngokungenamkhethe nenkcazo yeepropathi eziphambili ngokwenza izinto ezinobunzima kwaye ngenxa yoko, imifanekiso yenziwe ngokucacileyo azinakuchazwa kwiifoto zokwenyani.
Inguqulelo yokuqala yetekhnoloji ye-StyleGAN (ekhutshwe ngo-2019), elandelwa luhlobo oluphuculweyo lwe-StyleGAN2 ngo-2020, ephucula umgangatho womfanekiso kunye nokususa ezinye izinto zakudala. Kwangelo xesha, le nkqubo yahlala ihleli, oko kukuthi, ayivumeli oopopayi bokwenyani okanye intshukumo yobuso. Xa usenza i-StyleGAN3, eyona njongo iphambili yayikukuhlengahlengisa itekhnoloji ukuze isetyenziswe kupopayi nakwividiyo.
Uhlobo lweGAN3 lusebenzisa uyilo loyilo olungasasebenziyoI-ay ibonelela ngemeko zoqeqesho zenethiwekhi ezintsha kwaye ikwabandakanya izixhobo ezitsha zokubonisa ukusebenzisana (visualizer.py), uhlalutyo (avg_spectra.py) kunye nokuveliswa kwevidiyo (gen_video.py). Ukuphunyezwa kukwehlisa ukusetyenziswa kwememori kwaye kukhawulezise inkqubo yokufunda.
Olona phawu lubalulekileyo loyilo lwe-StyleGAN3 yayilutshintsho kutoliko lwayo yonke imiqondiso kwinethiwekhi ye-neural ngohlobo lweenkqubo eziqhubekayo, ezenza ukuba kube lula ukulawula izikhundla ezinxulumene nokwenza iinxalenye, ezingabotshwanga kulungelelwaniso olupheleleyo lweepikseli ezizodwa umfanekiso, kodwa ulungelelaniswe kumphezulu wezinto ezimelweyo.
Ngexesha kwi-StyleGAN kunye ne-StyleGAN2, ukufota kwiiphikseli ngexesha lokwakha kubangele imiba ngonikezelo olunamandlaUmzekelo, xa umfanekiso bewuhamba, bekukho ukungangqinelani kweenkcukacha ezincinci, ezinje ngemibimbi kunye neenwele, ezibonakala ngathi zihamba ngokwahlukeneyo nomfanekiso wobuso, ukongeza kuloo kwi-StyleGAN3 ezi ngxaki zisonjululwe kwaye ubuchwepheshe buye ilungele ukuveliswa kwevidiyo.
Ekugqibeleni, kufanelekile ukuba ukhankanye isibhengezo indalo eyenziwe yi-NVIDIA kunye neMicrosoft yemodeli enkulu yolwimi ye-MT-NLG esekwe kuthungelwano olunzulu lwe-neural kunye notshintsho «loyilo.
Imodeli igubungela iiparamitha ezingama-530 ezigidigidi kunye nephuli ye-4480 GPUs yasetyenziswa yoqeqesho (iiseva ezingama-560 DGX A100 ezinee-8 A100 GPUs ezingama-80 GB inye). Imimandla yokusetyenziswa kwemodeli ibizwa ngokuba kukusombulula ingxaki kulwimi lwendalo, njengokuqikelela ukugqitywa kwesivakalisi esingagqitywanga, ukuphendula imibuzo, ukuqonda ukuqonda, ukwenza izigqibo ngolwimi lwendalo, kunye nokuhlalutya ukungaqondakali kwentsingiselo yamagama.
Ukuba unomdla wokwazi okungakumbi ngayo, ungakhangela iinkcukacha ze-StyleGAN3 Kule khonkco ilandelayo.