Ngokuvamile, amamodeli akhiqizayo ajulile njengama-GAN, ama-VAE, namamodeli we-autoregressive aphatha izinkinga zokuhlanganiswa kwezithombe.
Uma kubhekwa ikhwalithi ephezulu yedatha abayidalayo, amanethiwekhi ezitha ezikhiqizayo (ama-GAN) athole ukunakwa okuningi eminyakeni yakamuva.
Amamodeli okusabalalisa angomunye umkhakha wocwaningo othakazelisayo osezisungule wona. Izinkambu zesithombe, ividiyo, kanye nokukhiqizwa kwezwi zombili zithole ukusetshenziswa okubanzi kwazo zombili.
Amamodeli okusabalalisa ngokumelene nama-GAN: Iyiphi Ekhiqiza Imiphumela Engcono? Ngokwemvelo, lokhu kuye kwaholela engxoxweni eqhubekayo.
Ezakhiweni zekhompyutha ezaziwa nge-GAN, ezimbili amanethiwekhi we-neural ziyalwa ukuze kukhiqizwe izimo ezisanda kuhlanganiswa zedatha engadlulela kudatha yangempela.
Amamodeli okusabalalisa aya ngokuya aduma njengoba enikeza ukuzinza kokuqeqeshwa kanye nemiphumela ephezulu yokukhiqiza umculo nezithombe.
Lesi sihloko sizodlula imodeli yokusabalalisa kanye nama-GAN ngokuningiliziwe, nokuthi ahluke kanjani komunye nomunye kanye nezinye izinto ezimbalwa.
Ngakho-ke, yini i-Generative Adversarial Networks?
Ukuze kudalwe okusha, izehlakalo zokwenziwa zedatha okungenzeka ukuthi kwenziwe iphutha ngedatha yangempela, amanethiwekhi akhiqizayo adversarial (GANs) asebenzisa amanethiwekhi emizwa amabili futhi awaxabanise (ngaleyo ndlela "ophikisanayo" egameni).
Asetshenziswa kakhulu ekukhulumeni, ekudaleni izithombe namavidiyo.
Inhloso ye-GAN wukudala idatha engatholwanga ngaphambilini kusuka kudathasethi ethile. Ukuzama ukuthola imodeli yokusabalalisa kwedatha okuyisisekelo kwangempela, okungaziwa kusuka kumasampuli, kwenza lokhu.
Okunye kushiwo, lawa manethiwekhi angamamodeli angacacile azama ukufunda ukusatshalaliswa kwezibalo okuthile.
Indlela eyasetshenziswa i-GAN ukuthola indlela yokufeza le nhloso kwakuyinoveli. Eqinisweni, bakhiqiza idatha ngokudlala igeyimu yabadlali ababili ukuze bathuthukise imodeli engacacile.
Okulandelayo kuchaza isakhiwo:
- Ubandlululo ozuza amandla okuhlukanisa phakathi kwedatha eyiqiniso nengamanga
- ijeneretha ecosha izindlela ezintsha zokwenza idatha ingakhohlisa umbandlululi.
Umbandlululi uzenza inethiwekhi ye-neural. Ngakho-ke, i-generator idinga ukudala isithombe esinekhwalithi ephezulu ukusikhohlisa.
Iqiniso lokuthi lawa majeneretha awaqeqeshwanga kusetshenziswa noma yikuphi ukusatshalaliswa kokuphumayo kuwumehluko obalulekile phakathi kwamamodeli eencoder ezenzakalelayo namanye amamodeli.
Kunezindlela ezimbili zokubola umsebenzi wokulahlekelwa wemodeli:
- ikhono lokulinganisa uma umbandlululi ebona kusengaphambili ngokunembile idatha yangempela
- idatha ekhiqiziwe ibikezelwa ngokunembile ngengxenye.
Ngokubandlulula okungcono kakhulu okungenzeka, lo msebenzi wokulahlekelwa ube usuncishiswa:
Ngakho-ke amamodeli ajwayelekile angacatshangwa njengamamodeli okunciphisa ibanga futhi, uma umbandlululi elungile, njengokunciphisa ukwehlukana phakathi kokusabalalisa kweqiniso nokukhiqiziwe.
Eqinisweni, ukuhlukana okuhlukene kungase kusetshenziswe futhi kuphumele ezindleleni ezihlukahlukene zokuqeqesha ze-GAN.
Amandla okufunda, ahlanganisa ukuhwebelana phakathi kwejeneretha kanye nomhlukanisi, kuyinselele ukulandela, naphezu kokuthi kulula ukulungisa umsebenzi wokulahlekelwa wama-GAN.
Azikho futhi iziqinisekiso zokuthi ukufunda kuzohlangana. Ngenxa yalokho, ukuqeqesha imodeli ye-GAN kunzima, njengoba kuvamile ukubhekana nezinkinga ezifana nama-gradient anyamalalayo kanye nokuwa kwemodi (uma kungekho ukuhlukahluka kumasampuli akhiqiziwe).
Manje, sekuyisikhathi samamodeli we-Diffusion
Inkinga yokuhlangana kokuqeqeshwa kwama-GAN iye yaxazululwa ngokuthuthukiswa kwamamodeli okusabalalisa.
Lawa mamodeli acabanga ukuthi inqubo yokusabalalisa ilingana nokulahlekelwa kolwazi okulethwa ukuphazamiseka okuqhubekayo komsindo (umsindo we-gaussian uyengezwa ngaso sonke isikhathi senqubo yokusabalalisa).
Inhloso yemodeli enjalo ukunquma ukuthi umsindo uthinta kanjani ulwazi olukhona kusampula, noma, ukukubeka ngenye indlela, ukuthi lungakanani ulwazi olulahlekile ngenxa yokusabalalisa.
Uma imodeli ingathola lokhu, kufanele ikwazi ukubuyisa isampula yoqobo futhi ihlehlise ukulahleka kolwazi okwenzekile.
Lokhu kufezwa ngemodeli yokusabalalisa i-denoising. Inqubo yokusabalalisa phambili kanye nenqubo yokuhlehla yokusabalalisa yenza lezi zinyathelo ezimbili.
Inqubo yokusabalalisa phambili ifaka kancane kancane ukungeza umsindo we-Gaussian (okungukuthi, inqubo yokusabalalisa) kuze kube yilapho idatha isingcoliswe ngokuphelele umsindo.
Inethiwekhi ye-neural kamuva iqeqeshwa kusetshenziswa indlela yokusabalalisa okuhlanekezelwe ukuze ifunde amathuba okusabalalisa anemibandela ukuze kuhlehliswe umsindo.
Lapha ungakwazi ukuqonda okwengeziwe mayelana imodeli yokusabalalisa.
Imodeli Yokusabalalisa vs ama-GAN
Njengemodeli yokusabalalisa, ama-GAN akhiqiza izithombe ngomsindo.
Imodeli yakhiwe inethiwekhi ye-neural generator, eqala ngomsindo wesimo esithile esishintshashintshayo esifundisayo, njengelebula yekilasi noma umbhalo wekhodi.
Umphumela kufanele ube into efana nesithombe esingokoqobo.
Ukuze udale izizukulwane zezithombe ze-photorealistic nezethembekile, sisebenzisa ama-GAN. Okubonwayo okungokoqobo nakakhulu kunama-GAN akhiqizwa kusetshenziswa amamodeli okusabalalisa.
Ngandlela thize, amamodeli okusabalalisa anembe kakhulu ekuchazeni amaqiniso.
Ngenkathi i-GAN ithatha njengomsindo ongahleliwe wokufakwayo noma okuguquguqukayo kwesimo sekilasi futhi ikhipha isampula yangempela, amamodeli okusabalalisa ngokuvamile awasheshi, ayaphindaphinda, futhi adinga ukuqondiswa okwengeziwe.
Asikho isikhala esikhulu sephutha uma i-denoising isetshenziswa ngokuphindaphindiwe ngomgomo wokubuyela esithombeni sokuqala kusuka emsindweni.
Indawo yokuhlola ngayinye idlula kuso sonke isigaba sokudalwa, futhi ngesinyathelo ngasinye, isithombe singase sizuze ulwazi olwengeziwe.
Isiphetho
Sengiphetha, Ngenxa yocwaningo olubalulekile olumbalwa olwanyatheliswa kuphela ngawo-2020 nango-2021, amamodeli okusabalalisa manje angakwazi ukudlula ama-GAN ngokuya ngokuhlanganisa izithombe.
Kulo nyaka, i-OpenAI yethulwe I-DALL-E2, imodeli yokukhiqiza isithombe evumela odokotela ukuthi basebenzise amamodeli okusabalalisa.
Nakuba ama-GAN esezingeni eliphezulu, izingqinamba zawo zikwenza kube inselele ukukala nokuwasebenzisa ezimweni ezintsha.
Ukuze kuzuzwe ikhwalithi yesampula efana ne-GAN usebenzisa amamodeli asekelwe okungenzeka, mningi umsebenzi owenziwe kuyo.
shiya impendulo