I-OpenAI Point E: Yenza i-3D point cloud ukusuka kumaza antsonkothileyo kwimizuzu kwi-GPU enye

Kwinqaku elitsha Point-E: Inkqubo yokuvelisa amafu e-3D ukusuka kwimiqondiso entsonkothileyo, iqela lophando le-OpenAI lazisa i-Point E, i-3D point cloud text text conditional synthesis system esebenzisa iimodeli zokusasazwa ukwenza iimilo ze-3D ezahlukeneyo nezintsonkothileyo eziqhutywa ngumbhalo onzima. iimpawu.kwimizuzu kwiGPU enye.
Ukusebenza okumangalisayo kweemodeli zanamhlanje zokuvelisa imifanekiso iye yavuselela uphando kwisizukulwana sezinto ezibhaliweyo ze-3D.Nangona kunjalo, ngokungafaniyo neemodeli ze-2D, ezinokuvelisa imveliso ngemizuzu okanye nakwimizuzwana, iimodeli zokuvelisa into zihlala zifuna iiyure ezininzi zomsebenzi we-GPU ukuvelisa isampulu enye.
Kwinqaku elitsha Point-E: Inkqubo yokuvelisa i-3D point clouds from complex signals, iqela lophando le-OpenAI libonisa i-Point · E, inkqubo ye-textual synthesis system ye-3D point clouds.Le ndlela intsha isebenzisa imodeli yosasazo ukwenza iimilo ze-3D ezahlukeneyo nezintsonkothileyo ukusuka kwimiqondiso yeteksti entsonkothileyo ngomzuzu nje okanye emibini kwiGPU enye.
Iqela ligxile kumceli mngeni wokuguqula isicatshulwa kwi-3D, ebaluleke kakhulu kwidemokhrasi yokudalwa komxholo we-3D kwizicelo zehlabathi zangempela ezivela kwinyani yenyani kunye nokudlala ukuya kuyilo lwamashishini.Iindlela ezikhoyo zokuguqula isicatshulwa kwi-3D ziwela kwiindidi ezimbini, nganye kuzo ineentsilelo zayo: 1) iimodeli ezivelisayo zingasetyenziselwa ukuvelisa iisampulu ngokufanelekileyo, kodwa azikwazi ukulinganisa ngokufanelekileyo kwiimpawu zesicatshulwa ezahlukeneyo kunye neziyinkimbinkimbi;I-2) imodeli yomfanekiso wombhalo oqeqeshwe kwangaphambili ukusingatha iimpawu eziyinkimbinkimbi kunye nezahlukeneyo zesicatshulwa, kodwa le ndlela ixakeke kakhulu kwaye imodeli inokubambelela ngokulula kwi-minima yendawo engahambelani nezinto ezinentsingiselo okanye ezihambelanayo ze-3D.
Ke ngoko, iqela liphonononge enye indlela ejolise ekudibaniseni amandla ezi ndlela zimbini zingasentla, kusetyenziswa imodeli yokusasazwa kwesicatshulwa ukuya kumfanekiso oqeqeshwe kwisethi enkulu yeebhangqa zomfanekiso wesicatshulwa (ukuvumela ukuba uphathe imiqondiso eyahlukeneyo kunye neyinkimbinkimbi) kunye imodeli ye-3D yokusasazwa komfanekiso oqeqeshwe kwiseti encinci yombhalo-mfanekiso ngababini.umfanekiso-3D pair dataset.Imodeli yokubhaliweyo ukuya kumfanekiso kuqala iisampulu yomfanekiso wegalelo ukwenza umboniso omnye wokwenziwa, kwaye imodeli-to-3D imodeli yenza i-3D point cloud ngokusekelwe kumfanekiso okhethiweyo.
Isitaki somyalelo sisekwe kwisakhelo soveliso esicetywayo samva nje sokuvelisa ngokwemiqathango imifanekiso esuka kwisicatshulwa (Sohl-Dickstein et al., 2015; Ingoma & Ermon, 2020b; Ho et al., 2020).Basebenzisa imodeli ye-GLIDE ene-3 yeebhiliyoni zeeparamitha ze-GLIDE (uNichol et al., 2021), ilungiswe kakuhle kwiimodeli ezinikezelweyo ze-3D, njengemodeli yabo yokuguqula umbhalo ukuya kumfanekiso, kunye neseti yeemodeli zokusasaza ezivelisa amafu e-RGB njengezabo. imodeli yenguqu.imifanekiso kumfanekiso.Iimodeli ze3D.
Ngelixa umsebenzi wangaphambili wasebenzisa i-architectures ye-3D ukucubungula amafu e-point, abaphandi basebenzisa imodeli elula esekelwe kwi-transducer (Vaswani et al., 2017) ukuphucula ukusebenza kakuhle.Kwimodeli yolwakhiwo lwabo lolwakhiwo, imifanekiso yelifu iqala yondliwa kwimodeli eqeqeshwe kwangaphambili yeViT-L/14 CLIP kwaye emva koko i-meshes ephumayo ifakwa kwisiguquli njengabamakishi.
Kuphononongo lwabo lobuchule, iqela lithelekisa indlela ecetywayo yePoint · E kunye nezinye iimodeli ze-3D eziveliswayo kumanqaku emiqondiso evela kubhaqo lwento yeCOCO, ulwahlulo, kunye neeseti zedatha.Iziphumo ziqinisekisa ukuba i-Point·E iyakwazi ukuvelisa iimilo ze-3D ezahlukahlukeneyo nezintsonkothileyo ukusuka kwimiqondiso yeteksti entsonkothileyo kunye nokukhawulezisa ixesha lokuqikelelwa ngomyalelo omnye ukuya kwezimbini zobukhulu.Iqela lithemba ukuba umsebenzi wabo uya kukhuthaza uphando olongezelelweyo kwi-3D text synthesis.
Imodeli yosasazo yelifu eqeqeshelwe kwangaphambili kunye nekhowudi yovavanyo iyafumaneka kwiGitHub yeprojekthi.Uxwebhu lweNdawo-E: Inkqubo yokudala i-3D point clouds from complex tracks ikwi-arXiv.
Siyazi ukuba awufuni kuphoswa naziphi na iindaba okanye ukufunyaniswa kwenzululwazi.Bhalisela kwincwadana yethu yeVeki ye-Synced Global AI ukuze ufumane uhlaziyo lweveki lwe-AI.