NVIDIA yatulutsa kachidindo ka StyleGAN3, makina ophunzirira makina ozikidwa pa generative adversarial neural network (GAN) yomwe cholinga chake ndi kupanga zithunzi zenizeni za nkhope za anthu. Khodiyo imalembedwa mu Python pogwiritsa ntchito dongosolo la PyTorch ndipo imagawidwa pansi pa NVIDIA Source Code License, yomwe imayika zoletsa kugwiritsa ntchito malonda.
Zitsanzo zokonzedwa kale zophunzitsidwa pagulu la Flickr-Faces-HQ (FFHQ), lomwe limaphatikizapo zithunzi za PNG 70 zapamwamba (1024x1024) za nkhope za anthu, ziliponso kuti zitsitsidwe. Kuonjezera apo, pali zitsanzo zomangidwa pamaziko a AFHQv2 (zithunzi za nkhope za nyama) ndi Metfaces (zithunzi za nkhope za anthu kuchokera pazithunzi za zojambula zakale) zosonkhanitsa. Cholinga cha chitukuko chimakhala pa nkhope, koma dongosololi likhoza kuphunzitsidwa kupanga zinthu zilizonse, monga malo ndi magalimoto. Kuphatikiza apo, zida zimaperekedwa kuti mudziphunzitse nokha neural network pogwiritsa ntchito zithunzi zanu. Pamafunika mmodzi kapena angapo NVIDIA zithunzi khadi (Tesla V100 kapena A100 GPU analimbikitsa), osachepera 12 GB wa RAM, PyTorch 1.9 ndi CUDA 11.1+ toolkit. Kuti mudziwe mawonekedwe opangira a nkhope zomwe zimachokera, chowunikira chapadera chikupangidwa.
Dongosolo limakupatsani mwayi wopanga chithunzi cha nkhope yatsopano kutengera kutanthauzira kwa mawonekedwe a nkhope zingapo, kuphatikiza mawonekedwe awo, komanso kusintha chithunzi chomaliza kukhala zaka zofunikira, jenda, kutalika kwa tsitsi, kumwetulira, mawonekedwe amphuno, khungu, magalasi, ndi mbali ya zithunzi. Jenereta amawona chithunzicho ngati mndandanda wa masitayelo, amangolekanitsa tsatanetsatane wa mawonekedwe (mawanga, tsitsi, magalasi) kuchokera kuzinthu zomwe zimafanana kwambiri (mawonekedwe, jenda, kusintha kwa zaka) ndikukulolani kuti muwaphatikize mwanjira iliyonse ndikutsimikiza kwamphamvu. katundu kudzera weighting coefficients. Chotsatira chake, zithunzi zimapangidwa zomwe sizimasiyanitsa ndi zithunzi zenizeni.
Mtundu woyamba waukadaulo wa StyleGAN udasindikizidwa mu 2019, pambuyo pake kusinthidwa kwa StyleGAN2020 kudapangidwa mu 2, kulola kuwongolera kwazithunzi ndikuchotsa zinthu zakale. Panthawi imodzimodziyo, dongosololi linakhalabe lokhazikika, i.e. sanalole kukwaniritsa makanema ojambula ndi kayendedwe ka nkhope. Popanga StyleGAN3, cholinga chachikulu chinali kusintha ukadaulo kuti ugwiritse ntchito pa makanema ojambula ndi makanema.
StyleGAN3 imagwiritsa ntchito kamangidwe kazithunzi kokonzedwanso, kopanda mawu, ndipo ikupereka maphunziro atsopano a neural network. Zimaphatikizapo zida zatsopano zowonera molumikizana (visualizer.py), kusanthula (avg_spectra.py) ndi kupanga makanema (gen_video.py). Kukhazikitsa kumachepetsanso kukumbukira kukumbukira ndikufulumizitsa njira yophunzirira.
Chinthu chofunika kwambiri pa zomangamanga za StyleGAN3 chinali kusintha kwa kutanthauzira zizindikiro zonse mu neural network monga momwe zimakhalira nthawi zonse, zomwe zinapangitsa kuti, popanga magawo, athe kusintha malo omwe sanagwirizane ndi ma pixel amtundu uliwonse. chithunzicho, koma chokhazikika pamwamba pa zinthu zomwe zawonetsedwa. Mu StyleGAN ndi StyleGAN2, kumangirira ma pixel m'mibadwo kumabweretsa mavuto panthawi yomasulira kwamphamvu, mwachitsanzo, chithunzicho chikasuntha, panali kusagwirizana kwazinthu zazing'ono, monga makwinya ndi tsitsi, zomwe zinkawoneka kuti zikuyenda mosiyana ndi nkhope yonse. . Mu StyleGAN3, mavutowa amathetsedwa ndipo ukadaulo wakhala woyenera pakupanga makanema.
Kuphatikiza apo, titha kuzindikira kulengeza za kulengedwa kwa NVIDIA ndi Microsoft yachitsanzo chachikulu kwambiri cha chilankhulo cha MT-NLG kutengera neural network yozama yokhala ndi "transformer" yomanga. Mtunduwu umakhudza magawo mabiliyoni 530, ndipo gulu la 4480 GPUs (maseva 560 DGX A100 okhala ndi 8 A100 80GB GPU iliyonse) adagwiritsidwa ntchito pophunzitsa. Ntchito zachitsanzozi zikuphatikizapo kuthetsa mavuto okhudza chinenero chachibadwa, monga kulosera kukwaniritsidwa kwa ziganizo zomwe sizinamalizidwe, kuyankha mafunso, kumvetsetsa kuwerenga, kujambula chinenero chachibadwa, ndi kusokoneza tanthauzo la mawu.
Source: opennet.ru