Inethiwekhi ye-neural yeNvidia ijika imizobo elula ibe yimbonakalo yomhlaba entle

Ingxangxasi yomtshayi kunye nengxangxasi yomntu osempilweni

Sonke siyayazi indlela yokuzoba isikhova. Okokuqala kufuneka uzobe i-oval, emva koko enye isangqa, kwaye emva koko ufumana isikhova esihle. Ngokuqinisekileyo, oku kuhlekisayo, kwaye kudala kakhulu, kodwa iinjineli zeNvidia zazama ukwenza ingcamango ibe yinyani.

Uphuhliso olutsha, ebizwa ngokuba yiGauGAN, yenza imihlaba emihle ukusuka kwimizobo elula kakhulu (ilula ngokwenene - izangqa, imigca kunye nayo yonke into). Ewe, olu phuhliso lusekwe kubuchwephesha bale mihla - oko kukuthi i-generative adversarial neural networks.

I-GauGAN ikuvumela ukuba udale umhlaba onemibala ebonakalayo - kungekuphela nje ukuzonwabisa, kodwa nomsebenzi. Ngoko ke, abakhi bezakhiwo, abaqulunqi bomhlaba, abaphuhlisi bemidlalo - bonke banokufunda into eluncedo. Ubukrelekrele bokwenziwa ngokukhawuleza "buqonda" into umntu ayifunayo kwaye izalisekisa ingcamango yokuqala kunye nenani elikhulu leenkcukacha.

"Ukucinga ngoyilo kulula kakhulu usebenzisa i-GauGAN kuba ibrashi ehlakaniphile inokuphucula umzobo wokuqala ngemifanekiso esemgangathweni," utshilo omnye wabaphuhlisi beGauGAN.

Abasebenzisi besi sixhobo banokutshintsha ingcamango yokuqala, ukuguqula i-landscape okanye omnye umfanekiso, ukongeza isibhakabhaka, isanti, ulwandle, njl. Nokuba yintoni enqwenelwa yintliziyo yakho, kwaye ukongeza kuthatha nje imizuzwana embalwa.

Inethiwekhi ye-neural yaqeqeshwa kusetyenziswa ugcino lwedatha yezigidi zemifanekiso. Ndiyabulela oku, inkqubo inokuqonda oko umntu akufunayo kunye nendlela yokufezekisa oko akufunayo. Ngaphezu koko, inethiwekhi ye-neural ayilibali malunga neenkcukacha ezincinci. Ngoko ke, ukuba uzobe umzobo oqingqiweyo wechibi kunye nemithi ethile ecaleni kwayo, emva kokuba umhlaba uphila, zonke izinto ezikufutshane ziya kubonakala kwisibuko samanzi echibini.

Inkqubo inokuxelelwa ukuba yintoni umphezulu obonakalayo omele ube yiyo - unokugqunywa ngotshani, ikhephu, amanzi okanye isanti. Konke oku kunokuguqulwa ngesibini, ukwenzela ukuba ikhephu libe yintlabathi kwaye endaweni yenkqantosi egqunywe yikhephu, umzobi ufumana indawo eyintlango.

“Ifana nencwadi yemibala ekuxelela indawo ofanele uwubeke kuyo umthi, indawo elikuyo ilanga, nendawo esikuyo isibhakabhaka. Emva koko, emva komsebenzi wokuqala, inethiwekhi ye-neural ivuselela umfanekiso, yongeza iinkcukacha eziyimfuneko kunye nokwenziwa, kwaye izoba imbonakalo. “Konke oku kusekwe kwimifanekiso yokwenyani,” utsho omnye wabaphuhlisi.

Nangona inkqubo ingenalo ihlabathi lokwenyani "lokuqonda", inkqubo ivelisa imbonakalo yomhlaba enomtsalane. Oku kungenxa yokuba iinethiwekhi ezimbini ze-neural zisetyenziswa apha, ijenereyitha kunye nomcaluli. Umenzi wenza umfanekiso kwaye uwubonise kumcaluli. Yena, ngokusekelwe kwizigidi zemifanekiso ebonwe ngaphambili, ukhetha olona khetho luyinyani.

Kungenxa yoko le nto ijenereyitha “iyazi” apho imbonakalo ifanele ibe khona. Kuyafaneleka ukuba uqaphele ukuba isixhobo siguquguquka kakhulu kwaye sixhotyiswe ngenani elikhulu lezicwangciso. Ngoko ke, ngoncedo lwayo, unokupenda imifanekiso, ukulungelelanisa nesitayela somculi othile, okanye udibanise ngokukhawuleza ukuphuma kwelanga okanye ukutshona kwelanga.

Abaphuhlisi babanga ukuba inkqubo ayithathi nje imifanekiso ukusuka kwenye indawo, yongeze kunye kwaye ufumane isiphumo. Hayi, yonke imiphumo "yemifanekiso" yenziwe. Oko kukuthi, inethiwekhi ye-neural "yenza" njengegcisa lokwenyani (okanye nangcono).

Okwangoku, inkqubo ayifumaneki ngokukhululekileyo, kodwa kungekudala kuya kwenzeka ukuyizama. Oku kunokwenziwa kwiNkomfa yeTekhnoloji yeGPU 2019, eyenzekayo ngoku eCalifornia. Abo banethamsanqa abakwaziyo ukutyelela umboniso bangakwazi ukuvavanya i-GauGAN.

Amanethiwekhi e-Neural kudala afundiswa ukuba athathe inxaxheba kwinkqubo yokudala. Ngokomzekelo, kunyaka ophelileyo, abanye babo inokwenza iimodeli ze3D. Ukongeza, abaphuhlisi abavela kwi-DeepMind baqeqeshe inethiwekhi ye-neural ukuba iphinde iphinde iphinde iphinde iphinde yakhiwe iindawo ezinokwakheka kuka-ntathu kunye nezinto ezivela kwimizobo, iifoto kunye neziketshi. Ukuze wenze kwakhona umfanekiso olula, inethiwekhi ye-neural idinga umfanekiso omnye; ukwenza izinto ezinzima ngakumbi, kufuneka imifanekiso emihlanu "yoqeqesho".

Ngokuphathelele iGauGAN, esi sixhobo siya kufumana ngokucacileyo isicelo esifanelekileyo sorhwebo - iindawo ezininzi zoshishino kunye nesayensi zinesidingo seenkonzo ezinjalo.

umthombo: www.habr.com