Indlela Ikati Lisuse Ngayo Ukusabalalisa Okuzinzile (2023)

Kwenye yezindaba ezingalindelekile zokulungisa iphutha emlandweni we-AI, ikati lasendlini lisize onjiniyela ngokungazi ukuthi bahlonze ukuhlanekezelwa kwesikhala okubalulekile epayipini lokukhiqiza isithombe le-Stable Diffusion. Isigameko sango-2023 sibe yisibonelo esiyingqopha-mlando sokuthi okokufaka komhlaba wangempela okungabikezeleki kungadalula kanjani amaphutha izinkulungwane zamahora okuhlolwa okuhleliwe awaphuthelwa nhlobo.

Kwenzekeni Ngempela Ngekati kanye Nokusabalalisa Okuzinzile?

Ekuqaleni kuka-2023, unjiniyela wokufunda ngomshini osebenza ekhaya wabona okuthile okungajwayelekile. Ikati labo, selihambe linqamula ikhibhodi ngesikhathi sokuqeqeshwa kwe-Stable Diffusion, lethule uchungechunge lwezinhlamvu ezingenangqondo eqeqebeni elisheshayo. Esikhundleni sokukhiqiza okuphumayo okubolile noma ukuphonsa iphutha, imodeli ikhiqize uchungechunge lwezithombe ezine-artifact ebonakalayo engaguquki futhi ecaciswe kakhulu — iphethini ye-tessellation ephindaphindayo obekungafanele ibe khona uma kubhekwa okokufaka okwenziwa ngokushesha.

Lona bekungelona umsindo ongahleliwe. Iphethini yembule ukuchema okungakaze kubonwe ngaphambili ezendlalelo zokunakwa kwemodeli, ikakhulukazi endleleni ukwakheka kwe-U-Net okucubungule ngayo inhlanganisela yamathokheni athile awela ngaphandle kwemingcele evamile yolimi. Ukuhlanganiswa kwekhibhodi yekati kudale ngempumelelo ukwazisa okungekho mhloli ongumuntu oke wacabanga ukuyizama, edalula iphutha ekuhlanganisweni kombhalo wekhodi we-CLIP wemodeli okuthinte indlela ubudlelwano bendawo obakhiwe ngayo ngesikhathi senqubo yokukhipha umsindo.

Ithimba lonjiniyela lichithe amaviki alandelayo lilandelela i-artifact emuva kumsuka wayo: inkinga yokuzungeza iphuzu elintantayo kusihleli esicashile sokusabalalisa esibonakala kuphela ngaphansi kwezimo ezithile zonqenqema lwethokheni. Ukulungisa kuthuthukise ukuhambisana kwesithombe kuzo zonke izinhlobo zokwaziswa ngesilinganiso esilinganiselwe esingu-3-4%, okuyinzuzo enkulu ekusebenzeni kwe-AI okukhiqizayo.

Kungani Okokufaka Okungajwayelekile Kubamba Izimbungulu Ezigejwa Amaqembu E-QA?

Ukuhlola okuhleliwe kulandela ingqondo yomuntu. Onjiniyela babhala izimo zokuhlola ngokusekelwe ekuziphatheni komsebenzisi okulindelekile, izimo ezisemaphethelweni abangazicabanga, nezindlela zokwehluleka ezaziwayo eziphindaphindwayo zangaphambilini. Kodwa isofthiwe - ikakhulukazi amasistimu e-AI anezigidigidi zamapharamitha - aqukethe ukuqhuma okuhlangene kwezimo okungenzeka ukuthi alukho uhlaka lokuhlola olungamboza ngokugcwele.

"Izimbungulu eziyingozi kakhulu akuzona lezi ezifihle ikhodi ongayihlolanga. Yizona ezifihla ikhodi oyihlolile ngokuqagela okungalungile." — Lesi simiso, esiqondwe isikhathi eside kubunjiniyela besofthiwe bendabuko, siba bucayi kakhulu ezinhlelweni zokufunda zomshini lapho isikhala sokufaka singapheli.

Isigameko sekati siqinise lokho osonjiniyela besiphithiphithi abakwaziyo iminyaka eminingi: okokufaka okungahleliwe, okungaqageleki kuveza ubuthakathaka besistimu obungakwazi ukuhlolwa kwendlela. Kungumgomo ofanayo ngemuva kokuhlolwa kwe-fuzz, lapho idatha engalungile ngamabomu ifakwa kumasistimu ukuze kwembulwe ubungozi. Umehluko lapha ukuthi i-fuzzer yayinemilenze emine nomsila.

Kuvezeni Lokhu Mayelana Nezinselele Zokulungisa I-AI?

Ukulungisa iphutha kumamodeli akhiqizayo e-AI ahluke kakhulu ekususeni iphutha kusofthiwe yosiko. Uma uhlelo lokusebenza oluvamile lwehluleka, uthola ilogi yephutha, ukulandelela isitaki, indlela ephindaphindekayo. Uma imodeli ye-AI ikhiqiza okuphumayo okungalungile ngokucashile, ukwehluleka kungase kungabonakali izinyanga ngoba ayikho impendulo eyodwa "elungile" engaqhathaniswa nayo.

• Ukufiphala kwesikhala esicashile: Izethulo zangaphakathi kumamodeli okusabalalisa ziyaziwa ukuthi kunzima ukuzihumusha, okwenza kube nzima ukulandelela ama-artifact okukhiphayo emuva ekuhlulekeni kwekhompuyutha okuthile.
• Ukuzwela okusheshayo: Ukwehluka okuncane kokufakwayo kombhalo kungaveza okukhiphayo okuhluke kakhulu, okusho ukuthi iziphazamisi zingavela kuphela ngaphansi kwezimo ezincane nezingalindelekile.
• Ukuzithoba kokuhlola: Ngokungafani nemisebenzi yokuhlukanisa ngokunemba okungakaleka, ikhwalithi yokukhiqiza isithombe incike ngokwengxenye, evumela ukucekelwa phansi okucashile ukuthi kudlule ekuhloleni okuzenzakalelayo.
• Ukuncika kwe-Cascading: Iphutha elilodwa kusifaki khodi sombhalo lingasabalala ngendlela yokunaka okuphambene, isihleli se-denoising, kanye nesiqophi se-VAE, okwenza ukuhlaziya imbangela yezimpande kube nzima kakhulu.
• Ukubanjwa kwedatha yokuqeqesha: Ukwehlukanisa phakathi kweziphazamisi kumodeli yezakhiwo nokuchema okuzuzwe njengefa kusuka kudatha yokuqeqeshwa kudinga izifundo zokucophelela ezidla isikhathi nezimba eqolo.

Lesi Sehlo Sibe Nethonya Kanjani Imikhuba Yokuthuthukisa I-AI?

Indaba yokulungisa iphutha lekati, ngenkathi ihlekisa ngaphandle, ibangele izinguquko ezithile ezingokoqobo endleleni amaqembu e-AI afinyelela ngayo isiqiniseko sekhwalithi. Izinhlangano eziningi selokhu zandise izivumelwano zazo zokuhlola i-fuzz zamamodeli akhiqizayo, ikakhulukazi ezihlanganisa ukulandelana kwamathokheni okungahleliwe nokuphikisanayo ezilingisa okokufaka okungezona ezolimi. Amanye amaqembu manje asebenzisa okuzenzakalelayo "ukuhamba kwekhibhodi" njengengxenye yamapayipi awo okuhlanganiswa okuqhubekayo.

Isigameko siphinde savuselela intshisekelo kumathuluzi okuhumusha amamodeli okusabalalisa. Ukube i-artifact ebonakalayo ibingacacile kangako - ukushintsha kombala okucashile kune-tessellation egqamile - bekungenzeka ukuthi ayizange ibonwe unomphela. Lokhu kuye kwaphusha umphakathi ukuthi uthuthukise ukutholwa okuzenzakalelayo okudidayo okungcono kokuphumayo, amasistimu angamaka ukungahambi kahle kwezibalo ngisho noma izithombe ngazinye zibonakala zivamile nje.

Emaqenjini aphethe ukugeleza komsebenzi okuyinkimbinkimbi kuyo yonke intuthuko ye-AI, ukuphindwaphindwa kwemikhiqizo, nokuqinisekiswa kwekhwalithi, izehlakalo ezifana nalezi zigqamisa isidingo sokubonakala kokusebenza okumaphakathi. Uma isiphazamisi sinweba isishumeki sombhalo, isihleli, kanye nesikhiphi khodi, ukulandelela uphenyo kuwo wonke amathuluzi ahlakazekile kanye namashaneli okuxhumana anqanyuliwe kudala ungqimba lwaso lokungqubuzana.

Imibuzo Evame Ukubuzwa

Ingabe isigameko sokususa iphutha sekati i-Stable Diffusion kwaba umcimbi wangempela?

Indaba ewumgogodla isuselwe ku-akhawunti eyabiwe kabanzi evela kumphakathi wonjiniyela we-AI ngo-2023. Nakuba imininingwane ethile yenziwe inganekwane ngandlela thile ekulandiseni kabusha, isimo sobuchwepheshe esiyisisekelo - okokufaka kwekhibhodi okungahleliwe okuveza isiphazamisi sesikhala esicashile - sibhalwe kahle futhi sihambisana nezindlela zokwehluleka ezaziwayo ezakhiweni zemodeli yokusabalalisa. Ukutholwa kwephutha okufanayo kwenzeke kuwo wonke umlando wobunjiniyela besofthiwe.

Ingabe ukuhlolwa kwe-fuzz kungakwazi ukubamba iziphazamisi ngendlela ethembekile kumamodeli e-AI akhiqizayo?

Ukuhlola i-Fuzz kusebenza kahle ekubambeni izigaba ezithile zeziphazamisi, ikakhulukazi lezo ezihlobene nokuhlaziya okokufaka, amakesi onqenqema wamathokheni, kanye nezinkinga zokuzinza kwezinombolo. Kodwa-ke, akuyona inhlamvu yesiliva ye-AI ekhiqizayo. Ngenxa yokuthi lawa mamodeli akhiqiza okuphumayo okungaba khona kunokunquma, ukuchaza ukuthi yini ehlanganisa "ukwehluleka" ngesikhathi sokuhlolwa kwe-fuzz kudinga amasistimu okuthola okudidayo ayinkimbinkimbi kunokugomela okulula/kokwehluleka.

Ingabe amaqembu e-AI angochwepheshe akuphatha kanjani ukugeleza komsebenzi kokulungisa amaphutha kuwo wonke amasistimu ayinkimbinkimbi?

Amathimba amaningi e-AI asebekhulile athembele kwinhlanganisela yezinkundla zokulandelela ukuhlolwa, ukugawulwa kwemithi okumaphakathi, imibhalo yokuhlanganyela, nokuphathwa kwephrojekthi okuhlelekile. Inselele eyinhloko iwukugcina ukulandeleka - ukuxhuma i-artifact ethile ephumayo enguqulweni yemodeli, idatha yokuqeqeshwa, ama-hyperparameters, kanye nokuzinikela kwekhodi okukhiqizile. Amathimba ahlanganisa lokhu kugeleza komsebenzi abe amasistimu okusebenza ahlangene achitha isikhathi esincane kakhulu ekuxhumaneni kanye nesikhathi esengeziwe ekuxazululeni izinkinga zangempela.

Yenza kubelula Ubukimbi Bakho Bokusebenza

Kungakhathaliseki ukuthi ulungisa amamodeli e-AI noma uphatha noma yikuphi ukusebenza kwebhizinisi okuyinkimbinkimbi, amathuluzi ahlukene adala ukucabanga okuyiziqephu. I-Mewayz iletha amamojula ahlanganisiwe angama-207 ohlelweni lokusebenza lwebhizinisi olulodwa oluthenjwa abasebenzisi abangaphezu kuka-138,000 - okunikeza iqembu lakho ukubonakala okumaphakathi okudingekayo ukulandelela izinkinga emthonjeni wazo, ukuxhumanisa izimpendulo, nokuhamba ngokushesha. Qala isilingo sakho samahhala ku-app.mewayz.com futhi ubone ukuthi ukusebenza okuhlanganisiwe kunjani.