GufunguraAI Ingingo ya E: Kurema igicu cya 3D kuva kumurongo uhindagurika muminota mike kuri GPU imwe

Mu kiganiro gishya Ingingo-E: Sisitemu yo kubyara ibicu bya 3D biturutse ku bimenyetso bigoye, itsinda ry’ubushakashatsi rya OpenAI ryerekana Point E, uburyo bwa 3D point igicu cyanditse cyerekana uburyo bwa synthesis ikoresha uburyo bwo gukwirakwiza imiterere ya 3D itandukanye kandi igoye itwarwa ninyandiko igoye. ibimenyetso.mu minota kuri GPU imwe.
Imikorere itangaje yuburyo bugezweho bwo kwerekana amashusho yerekana amashusho yashishikarije ubushakashatsi mubyakozwe mubintu bya 3D.Ariko, bitandukanye na 2D moderi, ishobora kubyara umusaruro muminota cyangwa amasegonda, ibintu bitanga ibintu mubisanzwe bisaba amasaha menshi yakazi ka GPU kugirango bitange icyitegererezo kimwe.
Mu kiganiro gishya Ingingo-E: Sisitemu yo kubyara ibicu bya 3D biturutse ku bimenyetso bigoye, itsinda ry’ubushakashatsi rya OpenAI ryerekana Point · E, sisitemu yo guhuza inyandiko ya sisitemu ya 3D point ibicu.Ubu buryo bushya bukoresha uburyo bwo kwamamaza kugirango habeho imiterere itandukanye kandi igoye ya 3D uhereye kumyandikire igoye muminota umwe cyangwa ibiri kuri GPU imwe.
Iri tsinda ryibanze ku mbogamizi zo guhindura inyandiko kuri 3D, zikaba ari ingenzi cyane mu guharanira demokarasi igizwe n'ibikorwa bya 3D ku bikorwa bifatika ku isi uhereye ku bintu bifatika ndetse no gukina ndetse no gushushanya inganda.Uburyo buriho bwo guhindura inyandiko kuri 3D bugabanijwe mubyiciro bibiri, buri kimwekimwe gifite ibibi: 1) moderi yibyara umusaruro irashobora gukoreshwa mugutanga ingero neza, ariko ntishobora gupima neza kubimenyetso bitandukanye kandi bigoye;)
Kubwibyo, itsinda ryashakishije ubundi buryo bugamije guhuza imbaraga zuburyo bubiri bwavuzwe haruguru, hakoreshejwe uburyo bwo gukwirakwiza inyandiko-ku-shusho yatojwe ku gice kinini cy’inyandiko-shusho (yemerera gukora ibimenyetso bitandukanye kandi bigoye) kandi icyitegererezo cya 3D ikwirakwiza icyitegererezo cyatojwe kumurongo muto winyandiko-shusho.ishusho-3D couple dataset.Icyitegererezo-cy-ishusho icyitegererezo cyambere cyerekana icyitegererezo cyo kwinjiza kugirango habeho icyerekezo kimwe, kandi ishusho-kuri-3D ikora igicu cya point point ya 3D ishingiye kumashusho yatoranijwe.
Itegeko ribyara umusaruro rishingiye kumurongo uherutse gutangwa kugirango habeho gutanga amashusho bivuye mumyandiko (Sohl-Dickstein et al., 2015; Indirimbo & Ermon, 2020b; Ho et al., 2020).Bakoresha icyitegererezo cya GLIDE hamwe na miliyari 3 z'ibipimo bya GLIDE (Nichol et al., 2021), bahujwe neza na moderi ya 3D yatanzwe, nk'icyitegererezo cyabo cyo guhindura amashusho, hamwe na moderi yo gukwirakwiza ibyara ibicu bya RGB nk'ibyabo icyitegererezo.amashusho.Moderi ya 3D.
Mugihe imirimo yabanje yakoresheje imyubakire ya 3D mugutunganya ibicu, abashakashatsi bakoresheje moderi yoroshye ya transducer (Vaswani et al., 2017) kugirango bongere imikorere.Muburyo bwo gukwirakwiza moderi yububiko, ingingo yibicu ishusho yabanje kugaburirwa muburyo bwateguwe mbere ya ViT-L / 14 CLIP hanyuma ibisohoka bisohoka bigaburirwa mubihindura nkibimenyetso.
Mu bushakashatsi bwabo bufatika, itsinda ryagereranije uburyo bwateganijwe bwa Point · E nubundi buryo bwa 3D butanga umusaruro ku bimenyetso byerekana amanota yatanzwe na COCO, gutondekanya, hamwe namakuru yashyizweho umukono.Ibisubizo byemeza ko Ingingo · E ishoboye kubyara imiterere itandukanye kandi igoye ya 3D uhereye kubimenyetso byanditse kandi byihutisha igihe cyo gufata umwanzuro kuri kimwe kugeza kuri bibiri byerekana ubunini.Itsinda ryizera ko akazi kabo kazashishikarizwa gukora ubushakashatsi kuri synthesis ya 3D.
Ingingo yagenwe mbere yo gukwirakwiza igicu no gusuzuma kode iraboneka kuri GitHub y'umushinga.Inyandiko Ingingo-E: Sisitemu yo gukora ibicu bya 3D yibicu biva mubimenyetso bigoye biri kuri arXiv.
Turabizi ko udashaka kubura amakuru cyangwa kuvumbura siyanse.Kwiyandikisha kumakuru yacu azwi cyane Synced Global AI Icyumweru kugirango wakire amakuru ya AI buri cyumweru.