A cikin 'yan shekarun nan, ƙirar ƙira da ake kira "samfuran watsawa" sun zama sananne, kuma tare da kyakkyawan dalili.
Duniya ta ga irin nau'ikan yadawa suke iyawa, kamar fitattun GANs akan haɗin hoto, godiya ga wasu zaɓaɓɓun wallafe-wallafen ƙasa waɗanda aka buga kawai a cikin 2020s & 2021s.
Kwararru kwanan nan sun ga yadda ake amfani da samfuran watsawa a ciki DALL-E2, Tsarin ƙirƙirar hoto na OpenAI wanda aka buga a watan da ya gabata.
Yawancin ma'aikatan Koyon Injin Babu shakka suna sha'awar ayyukan ciki na Model Diffusion sakamakon ci gaban da suka samu na kwanan nan.
A cikin wannan post ɗin, za mu kalli ƙa'idodin ƙa'idodin Tsarin Yadawa, ƙirar su, fa'idodin su, da ƙari mai yawa. Mu je.
Menene samfurin Yaduwa?
Bari mu fara da gano dalilin da yasa ake kiran wannan ƙirar azaman ƙirar yaduwa.
Kalmar da ke da alaƙa da thermodynamics a cikin azuzuwan physics ana kiranta watsawa. Tsarin ba ya cikin ma'auni idan akwai babban taro na abu, kamar ƙanshi, a wuri ɗaya.
Dole ne yaduwa ya faru don tsarin ya shiga daidaito. Kwayoyin kamshi suna yaduwa a cikin tsarin daga wani yanki mafi girma, yana sa tsarin ya zama daidai.
Komai daga ƙarshe ya zama kamanni saboda yaduwa.
Samfuran watsawa suna da kwarin gwiwa ta wannan yanayin yanayin ma'aunin ma'aunin zafi da sanyio. Samfuran watsawa suna amfani da sarkar Markov, wanda jerin masu canji ne inda ƙimar kowane maɓalli ya dogara da yanayin abin da ya gabata.
Ɗaukar hoto, a jere muna ƙara ƙayyadaddun amo gareshi a duk lokacin watsawar gaba.
Bayan adana hoton amo, za mu ci gaba da ƙirƙirar hoto na gaba a cikin jerin ta hanyar gabatar da ƙarin amo.
Sau da yawa, ana yin wannan hanya. Hoton amo mai tsabta yana haifar da maimaita wannan hanyar sau ƴan lokuta.
Ta yaya za mu iya ƙirƙirar hoto daga wannan ɗimbin hoto?
Ana juya tsarin yaduwa ta amfani da a neural network. Ana amfani da cibiyoyin sadarwa iri ɗaya da ma'auni iri ɗaya a cikin tsarin watsawa na baya don ƙirƙirar hoto daga t zuwa t-1.
Maimakon barin cibiyar sadarwa ta hango hoton, mutum na iya ƙoƙarin yin hasashen hayaniya a kowane mataki, wanda dole ne a cire shi daga hoton, don ƙara sauƙaƙe aikin.
A kowane yanayi, da ƙirar hanyar sadarwar jijiyoyi dole ne a zaba ta hanyar da ke kula da girman bayanai.
Zurfafa nutsewa cikin Samfurin Yaduwa
Abubuwan da ke cikin samfurin yaduwa sune tsari na gaba (wanda kuma aka sani da tsarin watsawa), wanda ake yin sautin datum (sau da yawa hoto) a hankali a hankali, da kuma wani tsari na baya (wanda aka sani da tsarin reverse diffusion), wanda a cikin amo yake. mayar da baya zuwa samfurin daga manufa rarraba.
Lokacin da matakin ƙara ya yi ƙasa sosai, ana iya amfani da Gaussians na sharaɗi don kafa sarƙar sarƙoƙi a cikin tsarin gaba. Sauƙaƙan ƙayyadaddun tsarin gaba yana haifar da haɗa wannan ilimin tare da zato Markov:
q (x1:T | x0): = YT t=1 q (xt|xt-1), q (xt|xt-1): = N (xt; p 1 - βtxt-1, βtI)
nan biyar….T shine jadawali bambance-bambance (ko koya ko gyarawa) wanda ke tabbatarwa, don isasshe babban T, cewa xT kusan Gaussian isotropic ne.
Akasin tsari shine inda sihirin samfurin yaduwa ya faru. Samfurin ya koyi jujjuya wannan tsarin watsawa yayin horo don samar da sabbin bayanai. Samfurin ya koyi rarraba haɗin gwiwa kamar yadda (x0:T) sakamakon farawa tare da tsantsar amo na Gaussian
(xT):=N (xT, 0, I).
pθ(x0:T):= p(xT) YT t=1 pθ(xt−1|xt), pθ(xt−1|xt):= N (xt−1; µθ (xt, t), Σθ( xt, t)))
inda aka gano sigogi masu dogaro da lokaci na Gaussian miƙa mulki. Musamman ma, lura da yadda tsarin Markov ya bayyana cewa rarrabawar canji da aka ba da baya ya dogara ne kawai akan lokutan da suka gabata (ko lokuta na gaba, dangane da yadda kuke kallonsa):
pθ(xt−1|xt):= N (xt-1; µθ (xt, t), Σθ(xt, t))
Koyarwar Model
Ana amfani da samfurin Markov mai juyowa wanda ke haɓaka yuwuwar bayanan horo don horar da ƙirar yaduwa. A zahiri, horo yana kwatankwacin rage bambance-bambancen babba akan yuwuwar log mara kyau.
E [- log pθ(x0)] ≤ Eq - log pθ(x0:T) q(x1:T |x0) = Eq - log p(xT) - X t≥1 log pθ(xt−1|xt) q (xt|xt-1) =: L
model
Yanzu muna buƙatar yanke shawarar yadda za mu aiwatar da Model Diffusion ɗinmu bayan kafa tushen ilimin lissafin aikin burin mu. Shawarar da kawai ake buƙata don aiwatar da gaba shine ƙayyadaddun jadawalin bambance-bambancen, wanda ƙimarsa yawanci ke tashi yayin aikin.
Muna la'akari sosai da yin amfani da sigar rarraba Gaussian da ƙirar ƙirar ƙira don juyawa hanya.
Yanayin ƙirar mu kawai shine cewa duka shigarwa da fitarwa suna da girma iri ɗaya. Wannan yana jadada ɗimbin 'yancin kai wanda Model Diffusion ke bayarwa.
A ƙasa, za mu shiga zurfin zurfi game da waɗannan zaɓuɓɓukan.
Tsarin Gabatarwa
Dole ne mu samar da jadawali bambance-bambance dangane da tsarin gaba. Mun sanya su musamman don zama masu dogaro da lokaci kuma mun yi watsi da yuwuwar za a iya koyan su. Jadawalin lokaci daga
β1 = 10-4 zuwa βT = 0.02.
Lt ya zama akai-akai dangane da saitin sigoginmu na koyo saboda ƙayyadaddun jadawalin bambance-bambancen, yana ba mu damar yin watsi da shi yayin horo ba tare da la'akari da takamaiman ƙimar da aka zaɓa ba.
Juya Tsari
Yanzu muna kan yanke shawarar da ake buƙata don ayyana tsarin baya. Ka tuna yadda muka bayyana baya Markov mika mulki a matsayin Gaussian:
pθ(xt−1|xt):= N (xt-1; µθ (xt, t), Σθ(xt, t))
Yanzu da muka gano nau'ikan aiki. Duk da cewa akwai ƙarin dabaru masu rikitarwa don daidaitawa, mun saita kawai
Σθ(xt, t) = σ 2 t I
σ 2 t = βt
Don sanya shi wata hanya, muna la'akari da Multivariate Gaussian don zama sakamakon daban-daban Gaussians tare da bambance-bambancen iri ɗaya, ƙimar bambance-bambancen da za ta iya canzawa akan lokaci. An saita waɗannan ɓangarorin don dacewa da jadawali na karkatar da tsari.
Sakamakon wannan sabon tsari, muna da:
pθ(xt−1|xt):= N (xt-1; µθ (xt, t), Σθ (xt, t)):=N (xt-1; µθ (xt, t), σ2 t I)
Wannan yana haifar da madadin aikin asarar da aka nuna a ƙasa, wanda marubutan suka gano don samar da ƙarin horo mai dacewa da sakamako mai kyau:
Sauki (θ): = Et, x0, h - θ
Har ila yau, marubutan sun zana alaƙa tsakanin wannan ƙirƙira na ƙirar yaduwa da ƙirar ƙira ta tushen Langevin. Kamar yadda yake tare da ci gaban kimiyyar lissafi mai zaman kansa na tushen igiyar ruwa da injiniyoyin ƙididdige matrix, waɗanda suka bayyana ƙayyadaddun ƙayyadaddun ƙira guda biyu na al'amura iri ɗaya, ya bayyana cewa Model Diffusion da Samfurin-Based Score na iya zama bangarori biyu na tsabar kuɗi ɗaya.
Cibiyar Gidan Hanya
Duk da cewa aikin asara na mu na nufin horar da abin ƙira θθ, Har yanzu ba mu yanke shawara game da gine-ginen wannan ƙirar ba. Ka tuna cewa samfurin kawai dole ne ya sami shigarwa iri ɗaya da girman fitarwa.
Ganin wannan ƙuntatawa, mai yiwuwa ba zato ba ne cewa ana amfani da gine-gine kamar U-Net akai-akai don ƙirƙirar samfuran watsa hotuna.
Ana yin sauye-sauye da yawa tare da hanyar tsarin juyawa yayin amfani da ci gaba da rarrabawar Gaussian. Ka tuna cewa makasudin tsarin baya shine ƙirƙirar hoto mai ƙima da ƙimar pixel lamba. Ƙayyade yiwuwar ƙima (log) ga kowane yuwuwar ƙimar pixel akan duk pixels ya zama dole.
Ana cim ma wannan ta hanyar keɓance keɓantaccen mai ƙididdigewa zuwa juzu'in juzu'i ta ƙarshe. kimanta damar wani hoto x0 da aka ba x1.
pθ(x0|x1) = YD i=1 Z δ+(xi 0) δ−(xi 0) N (x; µ i θ (x1, 1), σ2 1 ) dx
δ+(x) = ∞ idan x = 1 x + 1 255 idan x <1 δ−(x) = -∞ idan x = -1 x - 1 255 idan x > -1
inda babban rubutun na ke nuna cirewar haɗin kai ɗaya kuma D yana nuna adadin ma'auni a cikin bayanai.
Makasudin a wannan lokacin shine a tabbatar da yuwuwar kowane ƙimar lamba don takamaiman pixel da aka ba da rarrabuwar ƙima ga wannan pixel a cikin canjin lokaci. t=1.
Manufar Karshe
Mafi girman sakamako, a cewar masana kimiyya, sun fito ne daga yin hasashen sashin amo na hoto a wani lokaci. A ƙarshe, suna amfani da manufa mai zuwa:
Sauki (θ): = Et, x0, h - θ
A cikin hoton da ke tafe, ana siffanta hanyoyin horarwa da samfuran samfuran mu a takaice:
Fa'idodin Yaduwa Model
Kamar yadda aka riga aka nuna, adadin bincike akan samfuran watsawa ya ninka kwanan nan. Model Diffusion yanzu suna isar da ingancin hoto na zamani kuma ana samun wahayi ta hanyar ma'aunin ma'aunin zafi da sanyio.
Samfuran watsa shirye-shiryen suna ba da fa'idodi iri-iri ban da samun ingancin hoto mai yankan-baki, kamar rashin buƙatar horo na gaba.
Abubuwan da ke tattare da horon adawa an san su sosai, don haka galibi ya fi dacewa a zaɓi hanyoyin da ba na gaba ba tare da daidaitaccen aiki da ingantaccen horo.
Samfuran watsawa kuma suna ba da fa'idodin haɓakawa da daidaitawa dangane da tasirin horo.
Kodayake Model Diffusion ya bayyana yana haifar da sakamako da alama ba a cikin iska mai iska, tushen waɗannan sakamakon an ɗora shi ta hanyar ɗimbin shawarwari na lissafi masu tunani da ban sha'awa da dabara, kuma har yanzu ana haɓaka mafi kyawun ayyuka na masana'antu.
Kammalawa
A ƙarshe, masu bincike sun nuna ingantaccen binciken haɗin hoto ta amfani da samfuran yuwuwar watsawa, nau'in nau'ikan nau'ikan nau'ikan nau'ikan nau'ikan nau'ikan nau'ikan nau'ikan nau'ikan ra'ayi waɗanda ba su da ma'ana.
Sun cim ma manyan abubuwa godiya ga sakamakon da suka samu na fasaha na zamani da horon da ba na gaba ba da kuma ba da jarirai, ana iya tsammanin ƙarin ci gaba a cikin shekaru masu zuwa.
Musamman, an gano cewa samfuran watsawa suna da mahimmanci ga ayyukan ci-gaba na ƙira kamar DALL-E 2.
nan za ku iya samun damar cikakken bincike.
Leave a Reply