Indlela Ikati Elungisa Ngayo Ukusasazwa Okuzinzileyo (2023)

Kwelinye lawona mabali okulungisa obungalindelekanga kwimbali ye-AI, ikati yendlu ingaqondanga yanceda iinjineli ukuba zichonge eyona ndawo ifihlakeleyo yokugqwetheka kwesithuba kumbhobho wokuvelisa umfanekiso weStable Diffusion. Isehlo sika-2023 saba sisifundo esibalulekileyo kwindlela igalelo lehlabathi lokwenyani elingacingelwanga linokuthi liveze iziphene eziphoswa ngamawaka eeyure zovavanyo olucwangcisiweyo ngokupheleleyo.

Kwenzeke Ntoni Kanye Ngekati kunye noSahlulo oluZinzileyo?

Ebutsheni buka-2023, injineli yokufunda ngomatshini esebenza ekhaya yaqaphela into engaqhelekanga. Ikati yabo, emva kokuba behambe ngaphaya kwebhodi yezitshixo ngexesha loqeqesho oluZinzileyo lweDiffusion, bazisa umtya woonobumba abangekho ngqiqweni kwibhetshi ekhawulezayo. Endaweni yokuvelisa iziphumo ezigqwethekileyo okanye ukuphosa impazamo, imodeli ivelise uthotho lwemifanekiso ene-artifact ebonwayo engaguqukiyo necacileyo kakhulu - ipateni ephinda-phindayo yetessellation ebingamelanga ukuba ikhona xa kunikwa amagalelo akhawulezileyo.

Le ibingengxolo ingaqhelekanga. Ipatheni iveze i-bias engazange ibonwe ngaphambili kwimodeli yoqwalaselo olunqamlezayo, ngokukodwa kwindlela ukwakhiwa kwe-U-Net kuqhutywe ngayo imidibaniso ethile yomqondiso eyawa ngaphandle kwemida yolwimi eqhelekileyo. Ukudityaniswa kwebhodi yezitshixo zekati kwenze ngokufanelekileyo isilumkiso somchasi ekungekho mvavanyi ungumntu okhe wacinga ukuzama, eveza isiphene kumzekelo we-CLIP wodibaniso lwe-encoder text oluchaphazele indlela ubudlelwane besithuba obabalwa ngayo ngexesha lenkqubo yokukhupha i-denoising.

Iqela leenjineli lichithe iiveki ezilandelayo likhangela i-artifact emva konobangela wayo: umba ojikelezayo wenqaku elijikelezayo kumcwangcisi ofihlakeleyo wokusasazwa obonakala kuphela phantsi kweemeko ezithile zomphetho wophawu. Ulungiso luphucule ukuhambelana komfanekiso kuzo zonke iintlobo ezikhawulezayo ngoqikelelo lwe-3-4%, inzuzo ebalulekileyo ekusebenzeni kwe-AI.

Kutheni amagalelo aNgaqhelekanga eBambisa iiBugs eziphoswa ngamaqela e-QA?

Uvavanyo olucwangcisiweyo lulandela ingqiqo yomntu. Iinjineli zibhala iimeko zovavanyo ngokusekwe kwindlela yokuziphatha elindelekileyo yomsebenzisi, iimeko ezinokuthi zicinge, kunye neendlela ezaziwayo zokungaphumeleli ukusuka kuphindaphindo lwangaphambili. Kodwa isoftware - ngakumbi iinkqubo ze-AI ezineebhiliyoni zeeparameters - iqulethe uqhushumbo oludityanisiweyo lwamazwe anokwenzeka ukuba akukho sikhokelo sovavanyo sinokugubungela ngokupheleleyo.

"Ezona zinambuzane ziyingozi ayizizo ezizifihle kwikhowudi ongakhange uyivavanye. Zizo zifihla ikhowudi oyivavanye ngeengcinga ezingalunganga." — Lo mgaqo, kudala uqondwa kubunjineli besoftware yemveli, uba ngobaluleke ngakumbi kwiinkqubo zokufunda koomatshini apho indawo yokufaka ingenasiphelo.

Isehlo sekati sibethelele into eyaziwa ngabasebenzi bobunjineli iminyaka emininzi: amagalelo angenamkhethe, angalindelekanga atyhila ubuthathaka benkqubo obungenako ukuvavanywa ngendlela. Ikwangumgaqo ofanayo emva kovavanyo lwe-fuzz, apho idatha engalunganga ngabom ifakwa kwiisistim ukutyhila ubuthathaka. Umahluko apha yayikukuba ifuzi yayinemilenze emine nomsila.

Idiza Ntoni Le Ngemingeni ye-AI yokulungisa iimpazamo?

Ulungiso lweempazamo oluvelisa iimodeli ze-AI zahluke kakhulu kwisoftware yakudala yokulungisa ingxaki. Xa isicelo esiqhelekileyo singaphumeleli, ufumana i-log yemposiso, umkhondo wokupakisha, indlela yokuvelisa. Xa imodeli ye-AI ivelisa iziphumo ezingalunganga ngokufihlakeleyo, ukusilela kunokungaqapheli iinyanga kuba akukho mpendulo "echanekileyo" enye yokuthelekisa ngokuchaseneyo.

• Isithuba esifipheleyo sokungafihli: Umelo lwangaphakathi kwimifuziselo yosasazo luyaziwa ngokuba nzima ukulitolika, nto leyo eyenza kube nzima ukukhangela imveliso yobugcisa emva kokusilela okuthe ngqo kokubala.
• Uvakalelo olukhawulezileyo: Iiyantlukwano ezincinci kwigalelo lombhalo zinokuvelisa iziphumo ezahlukeneyo, okuthetha ukuba iibugs zinokuvela kuphela phantsi kweemeko ezimxinwa nezingalindelekanga.
• I-evaluation subjectivity: Ngokungafaniyo nemisebenzi yohlelo ngokuchaneka okunokumetwa, umgangatho wokwenziwa komfanekiso uxhomekeke ngokuyinxenye, uvumela ukuthotywa okufihlakeleyo ukuba kutyibilike kwiitshekhi ezizenzekelayo.
• Cascading dependents: Isiphene esinye kwi-encoder yokubhaliweyo sinokusasazeka ngokusebenzisa indlela yokuqaphela okunqamlezayo, i-denoising scheduler, kunye ne-VAE decoder, nto leyo eyenza uhlalutyo lwengcambu lubenzima kakhulu.
• Ukufakwa kwedatha yoqeqesho: Ukwahlula phakathi kweebhugi kuyilo loyilo kunye nokuthambekela okuzuzwe njengelifa kwidatha yoqeqesho kufuna uphononongo olucokisekileyo lokulahla oluthatha ixesha kunye nexabiso lekhompyuter.

Sizichaphazele Njani Ezi Sehlo kwiZenzo zoPhuhliso lwe-AI?

Ibali lokulungisa ikati, ngelixa lihlekisa kumphezulu, libangele utshintsho oluninzi kwindlela amaqela e-AI ajongana ngayo nesiqinisekiso somgangatho. Imibutho emininzi ukusukela oko yandise iiprothokholi zayo zovavanyo lwe-fuzz kwiimodeli ezivelisayo, ngakumbi ibandakanya ulandelelwano lwethokheni olungacwangciswanga kunye notshaba olulinganisa amagalelo angengolwimi. Amanye amaqela ngoku asebenzisa ukulinganisa "keyboard walk" ezenzekelayo njengenxalenye yemibhobho yokudibanisa eqhubekayo.

Isiganeko siphinde sahlaziya umdla kwizixhobo zokutolika kwiimodeli zosasazo. Ukuba i-artifact ebonakalayo ibingacacanga kangako-ukutshintsha kombala ofihlakeleyo endaweni ye-tessellation enesibindi-inokuba ibingabonwa ngokungenasiphelo. Oku kutyhalele abahlali ekuphuhliseni ubhaqo olungcono oluzenzekelayo oluzenzekelayo lweziphumo ezenziweyo, iinkqubo ezinokubonisa ukungahambi kakuhle kwamanani naxa imifanekiso nganye ibonakala iqhelekile.

Amaqela alawula ukuhamba komsebenzi okuntsonkothileyo kulo lonke uphuhliso lwe-AI, ukuphindaphindwa kwemveliso, kunye noqinisekiso lomgangatho, izehlo ezifana nezi ziqaqambisa imfuno yokubonakala kokusebenza okuphakathi. Xa ibug ijongisa ikhowudi yokubhaliweyo, umcwangcisi, kunye nesixhobo sokushicilela, ukulandelela uphando kuzo zonke izixhobo ezisasazekileyo kunye nemijelo yonxibelelwano eqhawulweyo yenza owayo umaleko wokukhuhlana.

Imibuzo Ebuzwa Rhoqo

Ingaba iStable Diffusion cat isehlo sokulungisa ingxaki sisehlo sokwenyani?

Ibali eliphambili lisekelwe kwi-akhawunti ekwabelwana ngayo ngokubanzi evela kuluntu lobunjineli be-AI ngo-2023. Nangona iinkcukacha ezithile ziye zayintsomi ngokuphindaphinda, imeko ephantsi yobugcisa - i-random keyboard input eveza i-bug yendawo efihliweyo - ibhalwe kakuhle kwaye ihambelana neendlela zokungaphumeleli ezaziwayo kwi-diffusion model architectures. Ukufunyaniswa ngengozi okufanayo kwenzeke kuyo yonke imbali yobunjineli besoftware.

Ngaba uvavanyo lwe-fuzz luyakwazi ukubamba ii-bugs kwiimodeli ze-AI?

Uvavanyo lwe-Fuzz luyasebenza ekubambiseni iindidi ezithile zeebugs, ngakumbi ezo zinxulumene nokwahlulahlula okokufaka, iimeko zomphetho we-tokenization, kunye nemiba yozinzo lwamanani. Nangona kunjalo, ayisiyombumbulu yesilivere ye-AI yokuvelisa. Ngenxa yokuba le mifuziselo ivelisa iziphumo ezinokwenzeka kunezingqinisiso, ukuchaza ukuba yintoni na eyenza "ukusilela" ngexesha lovavanyo lwe-fuzz kufuna iinkqubo ezintsonkothileyo zokubona okungaqhelekanga kunokupasa/ukungaphumeleli kweengxelo ezilula.

Amaqela achwephesha e-AI alawula njani ukuhanjiswa komsebenzi kwiinkqubo ezintsonkothileyo?

Uninzi lwamaqela e-AI aqolileyo axhomekeke kwindibaniselwano yamaqonga okulandelela umfuniselo, ukugawulwa kwemithi kwindawo esembindini, amaxwebhu entsebenziswano, kunye nolawulo lweprojekthi olucwangcisiweyo. Umceli mngeni ophambili kukugcina ukulandeleka - ukudibanisa i-artifact ethile yemveliso kwinguqulo yemodeli, idatha yoqeqesho, i-hyperparameters, kunye nekhowudi yokuzibophelela eyivelisayo. Amaqela ahlanganisa oku kuhamba komsebenzi kwiinkqubo zokusebenza ezidityanisiweyo achitha ixesha elincinci kakhulu kulungelelaniso kunye nexesha elingakumbi ekusombululeni ingxaki.

Yenza lula ukuntsonkotha kwakho kokuSebenza

Nokuba ulungisa iimodeli ze-AI okanye ulawula nayiphi na eminye imisebenzi entsonkothileyo yeshishini, izixhobo eziqhekekileyo zenza iingcinga eziqhekekayo. I-Mewayz izisa iimodyuli ezihlanganisiweyo ezingama-207 kwinkqubo yokusebenza yeshishini elinye ethenjwa ngabasebenzisi abangaphezu kwe-138,000 - inika iqela lakho ukubonakala okusembindini okufunekayo ukulandelela iingxaki kumthombo wabo, ukulungelelanisa iimpendulo, kunye nokuhamba ngokukhawuleza. Qala isilingo sakho sasimahla kwi-app.mewayz.com kwaye ubone ukuba imisebenzi emanyeneyo injani.