Mapopoma emuputi nemapopoma emunhu ane hutano
Tese tinoziva kudhirowa zizi. Kutanga iwe unofanirwa kudhirowa ovha, ipapo imwe denderedzwa, uye ipapo - iwe unowana gorgeous owl. Zvechokwadi, ichi chiseko, uye chekare, asi Nvidia mainjiniya akaedza kuita kuti fungidziro iitike.
, iyo inonzi GauGAN, inogadzira nzvimbo dzakanaka kubva kune dzakareruka sketches (yakareruka - madenderedzwa, mitsetse uye zvese). Ehe, budiriro iyi yakavakirwa pamatekinoroji emazuva ano - kureva, generative adversarial neural network.
GauGAN inokutendera kuti ugadzire nyika dzine mavara - uye kwete zvekunakidzwa chete, asiwo zvebasa. Saka, vavaki, vagadziri venzvimbo, vanogadzira mitambo - vese vanogona kudzidza chimwe chinhu chinobatsira. Artificial intelligence pakarepo "inonzwisisa" izvo munhu anoda uye anozadzisa iyo yekutanga pfungwa nehuwandu hukuru hweruzivo.
"Kufunga nezvekuvandudza dhizaini kuri nyore kwazvo nerubatsiro rweGauGAN, sezvo bhurashi rakangwara richikwanisa kutsigira sketch yekutanga nekuwedzera mifananidzo yemhando," akadaro mumwe mugadziri weGauGAN.
Vashandisi veichi chishandiso vanogona kushandura iyo yekutanga pfungwa, kugadzirisa mamiriro kana mumwe mufananidzo, kuwedzera denga, jecha, gungwa, nezvimwe. Zvese zvinoda moyo wako, uye kuwedzera kunotora mashoma mashoma.
Iyo neural network yakadzidziswa uchishandisa database yemamiriyoni emifananidzo. Nekuda kweizvi, iyo system inogona kunzwisisa izvo munhu anoda uye maitiro ekuita izvo zvavanoda. Uyezve, iyo neural network haikanganwi nezve zvidiki zvidiki. Saka, kana iwe ukaita schematically kudhirowa dziva nemimwe miti iri pedyo naro, ipapo mushure mokunge nyika yamutsidzirwa, zvinhu zvose zviri pedyo zvicharatidzwa mugirazi remvura yemvura.
Iwe unogona kutaurira sisitimu izvo zvinoonekwa zvinofanirwa kuve - zvinogona kuvharwa neuswa, chando, mvura kana jecha. Zvose izvi zvinogona kushandurwa muchipiri, kuitira kuti chando chive jecha uye panzvimbo yemarara echando, nyanzvi inowana nzvimbo yerenje.
βZvakafanana nebhuku remavara rinotaura pokuisa muti, zuva riri kupi, uye denga riripi. Zvino, mushure mebasa rekutanga, iyo neural network inosimudzira mufananidzo, inowedzera inodiwa ruzivo uye maumbirwo, inodhirowa ratidziro. Zvese izvi zvinobva pamifananidzo chaiyo,β anodaro mumwe wevagadziri.
Kunyangwe iyo sisitimu ichishaya "kunzwisisa" kwenyika chaiyo, iyo system inogadzira inoyevedza nzvimbo. Izvi zvinodaro nekuti maviri neural network anoshandiswa pano, jenareta uye sarura. Jenareta inogadzira mufananidzo uye unoiratidza kune musaruri. Iye, zvichibva pamamiriyoni emifananidzo yakamboonekwa, anosarudza sarudzo dzechokwadi.
Ndicho chikonzero jenareta "inoziva" panofanira kuva nemafungiro. Izvo zvakakosha kuti tizive kuti chishandiso chakanyanya kuchinjika uye chakashongedzerwa nehuwandu hukuru hwezvirongwa. Saka, nayo, iwe unogona kupenda, kugadzirisa kune chimiro cheimwe nyanzvi, kana kungotamba-tamba nekukasira kuwedzera kwekubuda kwezuva kana kuvira kwezuva.
Vagadziri vacho vanoti sisitimu haingotore mifananidzo kubva kune imwe nzvimbo, wedzera pamwe chete uye uwane mhedzisiro. Kwete, "mifananidzo" yese yakagamuchirwa inogadzirwa. Ndokunge, iyo neural network "inogadzira" senge muimbi chaiye (kana kunyange zvirinani).
Kusvika ikozvino, chirongwa chacho hachiwaniki pachena, asi munguva pfupi zvichange zvichibvira kuiedza mubasa. Izvi zvinogona kuitwa kuGPU Technology Conference 2019, iyo iri kuitika muCalifornia. Vane rombo rakanaka vakakwanisa kushanyira kuratidzwa vanogona kutoedza GauGAN.
Neural network yakadzidziswa kwenguva refu kutora chikamu mukuita kwekugadzira. Somuenzaniso, gore rakapera, vamwe vavo . Pamusoro pezvo, vanogadzira kubva kuDeepMind vakadzidzisa neural network kudzoreredza nzvimbo dzine mativi matatu uye zvinhu kubva kudhirowa, mafoto, uye zvikeche. Kuti udzore zvakare chimiro chakareruka, mufananidzo mumwe wakakwana kune neural network, kugadzira zvimwe zvinhu zvakaoma, mifananidzo mishanu inodiwa pa "kudzidziswa".
Kana iri GauGAN, chishandiso ichi chichawana zvakajeka kushandiswa kwekutengesa kwakakodzera - nzvimbo zhinji dzebhizinesi nesainzi dzinoda masevhisi akadaro.
Source: www.habr.com