I-NVIDIA ipapashe ikhowudi yemvelaphi ye-StyleGAN3, inkqubo yokufunda ngomatshini esekwe kwinethiwekhi ye-neural ye-adversarial neural (GAN) ejolise ekuhlanganiseni imifanekiso eyinyani yobuso babantu. Ikhowudi ibhalwe kwiPython isebenzisa isakhelo sePyTorch kwaye ihanjiswa phantsi kweLayisensi yeKhowudi yoMthombo weNVIDIA, ebeka izithintelo ekusebenziseni urhwebo.
Iimodeli ezisele zenziwe ngobuchule eziqeqeshwe kwingqokelela ye-Flickr-Faces-HQ (FFHQ), ebandakanya i-70 yamawaka ekhwalithi ephezulu (1024x1024) imifanekiso yePNG yobuso babantu, nayo iyafumaneka ukukhuphela. Ukongeza, kukho iimodeli ezakhiwe ngesiseko se-AFHQv2 (iifoto zobuso bezilwanyana) kunye neeMetfaces (imifanekiso yobuso babantu ukusuka kwimifanekiso yokudweba kweklasikhi) ingqokelela. Ugxininiso lophuhliso lusebusweni, kodwa inkqubo inokuqeqeshwa ukuvelisa naziphi na izinto, ezifana neendawo kunye neemoto. Ukongeza, izixhobo zinikezelwe ukuziqeqesha ngokwakho inethiwekhi ye-neural usebenzisa ingqokelela yemifanekiso yakho. Ifuna enye okanye ngaphezulu ikhadi legraphic NVIDIA (Tesla V100 okanye A100 GPU kucetyiswa), ubuncinane 12 GB RAM, PyTorch 1.9 kunye CUDA 11.1+ toolkit. Ukumisela ubume bokwenziwa kobuso obubangelwayo, i-detector ekhethekileyo iyaphuhliswa.
Inkqubo ikuvumela ukuba wenze umfanekiso wobuso obutsha ngokusekwe ekudityanisweni kweempawu zobuso obuninzi, ukudibanisa iimpawu zabo, kunye nokulungelelanisa umfanekiso wokugqibela kwiminyaka efunekayo, isini, ubude beenwele, umlinganiswa woncumo, imilo yempumlo, umbala wolusu, iiglasi, kunye ne-engile yefoto. Ijeneretha ithatha umfanekiso njengengqokelela yesitayile, izahlula ngokuzenzekelayo iinkcukacha zeempawu (amafreckles, iinwele, iiglasi) ukusuka kwiimpawu eziqhelekileyo zomgangatho ophezulu (indawo, isini, utshintsho lweminyaka) kwaye ikuvumela ukuba udibanise kuyo nayiphi na ifom kunye nokuzimisela kokulawula. iipropati ngokusebenzisa i-coefficients yokulinganisa. Ngenxa yoko, kwenziwa imifanekiso engabonakaliyo kwiifoto zangempela.
Inguqulelo yokuqala yetekhnoloji ye-StyleGAN yapapashwa ngo-2019, emva koko kwacetywa ushicilelo oluphuculweyo lwe-StyleGAN2020 ngo-2, luvumela ukuphuculwa komgangatho wemifanekiso kunye nokuphelisa ezinye izinto zakudala. Ngelo xesha, inkqubo yahlala i-static, okt. khange ivumele ufezekiso lwepopayi oluyinyani kunye nentshukumo yobuso. Xa kuphuhliswa i-StyleGAN3, eyona njongo iphambili yayikukuqhelanisa itekhnoloji ukuze isetyenziswe kuopopayi kunye nevidiyo.
I-StyleGAN3 isebenzisa uyilo oluyilwe ngokutsha loyilo lwemifanekiso, simahla, kwaye iphakamisa iimeko ezintsha zoqeqesho kwinethiwekhi ye-neural. Iquka izinto ezintsha eziluncedo zokubonisa okusebenzisanayo (visualizer.py), uhlalutyo (avg_spectra.py) kunye nokuveliswa kwevidiyo (gen_video.py). Ukuphunyezwa kwakhona kunciphisa ukusetyenziswa kwememori kunye nokukhawuleza inkqubo yokufunda.
Uphawu oluphambili loyilo lweStyleGAN3 yayiyinguqulelo yokutolika yonke imiqondiso kuthungelwano lwe-neural ngendlela yeenkqubo eziqhubekayo, ezenza ukuba kwenzeke, xa kusenziwa iinxalenye, ukuguqula izikhundla ezihambelanayo ezingabotshelelwanga kulungelelwaniso olupheleleyo lweepixels ezizimeleyo. umfanekiso, kodwa umiliselwe kumphezulu wezinto ezibonisiweyo. KwiSitayileGAN kunye neSitayileGAN2, ukubophelela kwiipikseli ngexesha lesizukulwana kukhokelele kwiingxaki ngexesha lonikezelo oluguquguqukayo, umzekelo, xa umfanekiso ushukumayo, bekukho ukungangqinelani kweenkcukacha ezincinci, ezinje ngemibimbi kunye neenwele, ezibonakala zihamba ngokwahlukileyo kubuso bonke. . Kwi-StyleGAN3, ezi ngxaki zisonjululwe kwaye itekhnoloji ilungele ukuveliswa kwevidiyo.
Ukongeza, sinokuqaphela ukubhengezwa kokudalwa kwe-NVIDIA kunye neMicrosoft yemodeli enkulu yolwimi lweMT-NLG esekelwe kuthungelwano olunzulu lwe-neural kunye ne-architecture "transformer". Imodeli igubungela i-530 yeebhiliyoni zeeparamitha, kunye neqela le-4480 GPUs (iiseva ze-560 DGX A100 ezine-8 A100 80GB GPUs nganye) zasetyenziselwa uqeqesho. Usetyenziso lwalo mzekelo lubandakanya ukusombulula iingxaki zokusetyenzwa kolwimi lwendalo, njengokuqikelela ukugqitywa kwezivakalisi ezingagqitywanga, ukuphendula imibuzo, ukufunda ukuqonda, ukuzoba intelekelelo kulwimi lwendalo, kunye nokungabambisani intsingiselo yamagama.
umthombo: opennet.ru