Munguva pfupi yapfuura NVIDIA yakaburitsa kodhi kodhi yeiyo StyleGAN3, muchina wekudzidza muchina unobva mukugadzira yakashata neural network (GAN) yekugadzirisa mifananidzo chaiyo yezviso zvevanhu.
MuChimiroGAN3 zviripo kurodha akagadzirira-kushandisa-akadzidziswa mamodheru akadzidziswa muunganidzwa Flickr-Faces-HQ (FFHQ), iyo inosanganisira zviuru makumi manomwe zvemifananidzo yePNG yemhando yepamusoro yezviso zvevanhu (70 × 1024). Uye zvakare, pane mamodheru akavakirwa pahwaro hweAFHQv1024 (mifananidzo yezviso zvemhuka) uye Metface (mifananidzo yezviso zvevanhu kubva ku classical kupenda mifananidzo) kuunganidzwa.
Nezve StyleGAN3
Dhizaini inotarisa kumeso, asi iyo system inogona kudzidziswa kugadzira chero mhando yechinhu, kufanana nenzvimbo uye mota. Chii chimwe, zvishandiso zvinopihwa kuzvidzidzira-kudzidza kweiyo neural network uchishandisa yako wega mifananidzo yekuunganidza. Inoda imwe kana kupfuura makadhi emifananidzo eNVIDIA (Tesla V100 kana A100 GPUs yakakurudzirwa), ingangoita 12GB ye RAM, PyTorch 1.9, uye CUDA 11.1+ Toolkit. Kuti uone hunhu hwezviso zvakagamuchirwa, yakakosha detector iri kugadzirwa.
Iyo system inobvumidza kugadzira mufananidzo wechiso chitsva zvichibva pakapindirana kwezvinhu zvemeso akati wandei, Kubatanidza yavo yehunhu maficha, mukuwedzera kuchinjisa iyo yekupedzisira mufananidzo kune inodikanwa zera, jenda, kureba kwebvudzi, kunyemwerera hunhu, mhino chimiro, ganda ganda, magirazi, mifananidzo yemifananidzo.
Jenareta inobata iyo mufananidzo sekuunganidzwa kwemastaera, otomatiki anoparadzanisa hunhu ruzivo (mabara, vhudzi, magirazi) ehunhu hwepamusoro-soro (chimiro, murume kana murume, shanduko dzine chekuita nezera) uye zvinobvumidza kuzvisanganisa zvisina tsananguro nedudziro yezvivakwa zvakakurisa kuburikidza nekuyera zvinhu uye kuti semhedzisiro, mifananidzo inogadzirwa kuti ivo zviri pachena kuti hazvizivikanwe kubva pamifananidzo chaiyo.
Yekutanga vhezheni yeStyleGAN tekinoroji (yakaburitswa mu2019), inoteverwa neshanduro yakavandudzwa yeStyleGAN2 muna 2020, iyo inovandudza mhando yemifananidzo uye inobvisa zvimwe zvigadzirwa. Panguva imwecheteyo, iyo sisitimu yakaramba yakamira, ndiko kuti, haina kubvumira chaiwo maficha kana kumeso kwechiso. Paunenge uchigadzira StyleGAN3, chinangwa chikuru chaive chekugadzirisa tekinoroji kuti ishandiswe mumifananidzo uye vhidhiyo.
StyleGAN3 inoshandisa yakagadziriswazve isiri-yekufungidzira yekuvakisa dhizainiay inopa new neural network yekudzidzisa zviitiko uye zvakare inosanganisira zvishandiso zvitsva zvekudyidzana visualization (visualizer.py), kuongorora (avg_spectra.py) uye vhidhiyo chizvarwa (gen_video.py). Iyo yekumisikidza inoderedzawo ndangariro kushandiswa uye inomhanyisa maitiro ekudzidza.
Chinhu chakakosha cheiyo DhizainaGAN3 dhizaini kwaive kushandurwa kududzirwa kwemasaini ese ari munzira yehurongwa hwekuenderera, izvo zvakaita kuti zvikwanisike kushandisa zvigaro zvine hukama nekuumba zvikamu, zvisina kusungirirwa kumhedzisiro yemapikisheni ega ega mufananidzo, asi wakagadziriswa kumeso kwezvinhu zvakamiririrwa.
Nguva muChitaGAN uye StyleGAN2, kubaya mapikisheni panguva yekuvaka kwakakonzera nyaya nekushandura kwakasimbaSemuenzaniso, chifananidzo paichifamba, paive nekusawirirana kwezvinhu zvidiki, senge makwinya nemvere, izvo zvaiita kunge zvinofamba zvakasiyana nechero mufananidzo wechiso, kuwedzera kune iyo muStyleGAN3 matambudziko aya agadziriswa uye tekinoroji ive yakanyatsokodzera kugadzirwa kwevhidhiyo.
Pakupedzisira, zvakare zvakakodzera kutaura kuziviswa kwe iyo yakasikwa neNVIDIA neMicrosoft yemhando yakakura yemutauro weMT-NLG zvichibva pane yakadzika neural network ine »shanduko« dhizaini.
Iyo modhi inovhara 530 bhiriyoni paramita uye puru ye4480 GPUs yakashandiswa yekudzidzira (560 DGX A100 maseva ane 8 A100 GPUs e80 GB imwe neimwe). Idzo nzvimbo dzekushandiswa kwemuenzaniso dzinodaidzwa kunzi kugadzirisa dambudziko kugadzirisa mumutauro wechisikigo, sekufungidzira kupera kwechirevo chisina kupera, kupindura mibvunzo, kuverenga kuverenga, kugadzira mhedziso mumutauro wechisikigo, uye kuongorora kusanzwisisika kwezvinoreva mazwi.
Kana iwe uchifarira kuziva zvakawanda nezvazvo, unogona kutarisa ruzivo rwe StyleGAN3 Mune inotevera chinongedzo.