Mixup da CutMix Augmentation
Mixup da CutMix hanyoyin haɓaka bayanai ne waɗanda ke ƙirƙirar sabbin misalan horarwa ta hanyar haɗa hotuna biyu da alamun su.
Dubawa
Mixup linearly interpolates whole images and labels, while CutMix pastes a rectangular patch from one image onto another and mixes labels by patch area — both reduce overfitting and improve robustness.
Zurfafa nutsewa
Mixup (Zhang et al., 2017) ya samar da sabon samfurin kamar x̃ = λ·x_a + (1−λ) · x_b tare da lakabin ỹ gauraye ta λ guda ɗaya, inda aka zana λ daga rarrabawar Beta. Wannan yana ƙarfafa ƙirar don nuna hali tsakanin misalai, daidaita iyakokin yanke shawara da inganta daidaitawa. CutMix (Yun et al., 2019) a maimakon haka ya yanke yanki rectangular daga hoton B kuma ya manna shi akan hoton A; An saita ma'aunin lakabin da adadin pixels kowane hoto yana ba da gudummawa. Saboda CutMix yana kiyaye yankuna masu daidaituwa na gida (maimakon gaurayewar fatalwa), yana adana tsarin sararin samaniya mai amfani yayin da har yanzu yana tilasta samfurin halartar abubuwa da sassa da yawa. Dukansu fasahohin suna aiki ne azaman masu tsara tsarin lokaci mai ƙarfi, haɓaka daidaito akan ma'auni na ma'auni na ImageNet, kuma musamman inganta ƙarfi ga ɓarna da abubuwan gaba.
Fahimtar Fasaha
Duk hanyoyin biyu suna gyara maƙasudin asara, ba kawai shigarwar ba. Lakabin ya zama manufa mai laushi, gauraye, don haka asarar giciye-entropy shine haɗin λ-nauyin haɗe-haɗe na azuzuwan biyu - yadda ya kamata wani nau'i na alamar santsi da aka ɗaure da rabon hadawar pixel. A cikin CutMix, λ yana daidai da juzu'in pixels waɗanda ba a canza su ba, an ƙididdige su daga yankin akwatin da aka raba ta jimlar yanki na hoto, wanda ke kiyaye ƙimar alamar daidai da nawa ake iya gani na kowane hoto.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar Mixup da CutMix Augmentation
Augmentation na tushen cakuɗa yanzu ya zama daidaitaccen tsari a cikin ƙaƙƙarfan girke-girke na rarrabuwar hoto kuma yana tallafawa bututun horo na zamani don na'urorin canza hangen nesa, waɗanda galibi suna buƙatar daidaitawa sosai. Ana ci gaba da bincike akan bambance-bambancen sane-fadi (misali, sanya yankewa akan yankuna masu ba da labari), haɗuwa-matakin alama don masu canji, da kari zuwa bayanan sauti, rubutu, da bayanan 3D. Yi tsammanin dabarun haɗawa don zama ɗan lefa mai rahusa don haɓaka daidaito, daidaitawa, da ƙarfi yayin da gine-ginen ke ƙaruwa da yunwar bayanai.
Aiwatar da Gaskiyar Duniya
Horar da ImageNet classifiers tare da CutMix don haɓaka daidaito na sama-1 da haɓaka gano abubuwa.
Aiwatar da Mixup don inganta ƙirar ƙira don haka ƙwaƙƙwaran annabta sun fi dacewa da daidaito na gaskiya.
Ƙarfafa daidaita masu canjin hangen nesa (misali, DeiT) tare da haɗakar Mixup da CutMix don horar da kan iyakataccen bayanai.
Ƙarfafa ƙarfi ga ɓarnatar hoto da abubuwan da ba a rarrabawa a cikin tsarin hangen nesa mai mahimmancin aminci.
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Mixup and CutMix Augmentation quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Jagora na gaba
Gwaji-Lokacin Ƙarfafawa
Tambayoyin da ake yawan yi
What is Mixup and CutMix Augmentation?
Mixup da CutMix hanyoyin haɓaka bayanai ne waɗanda ke ƙirƙirar sabbin misalan horarwa ta hanyar haɗa hotuna biyu da alamun su. Haɗin kai tsaye yana haɗa hotuna da lakabi gabaɗaya, yayin da CutMix ke liƙa facin rectangular daga hoto ɗaya zuwa wani kuma yana haɗa lakabin ta wurin faci - duka suna rage wuce gona da iri da haɓaka ƙarfi.
Ta yaya Mixup ke ƙirƙirar sabon misalin horo?
Haɗa nau'ikan x̃ = λx_a + (1−λ)x_b kuma yana haɗa takubban tare da λ iri ɗaya da aka zana daga rarrabawar Beta.
Menene babban bambanci tsakanin CutMix da Mixup?
CutMix yana kwafin yanki mai rectangular daga hoto ɗaya zuwa wani, yana kiyaye abun ciki mai daidaituwa a cikin gida maimakon ɗaukar hotuna biyu tare.
A cikin CutMix, ta yaya ake ƙayyade nauyin haɗaɗɗiyar (lambda) don alamun?
Matsakaicin lakabin a cikin CutMix an saita shi ta yankin akwatin da aka liƙa dangane da ɗaukacin hoton, yana kiyaye alamun daidai da pixels na bayyane.
Bayan hoton shigarwa, menene Mixup da CutMix suke gyara?
Dukansu hanyoyin sun haɗu da alamun ma, suna samar da maƙasudai masu laushi waɗanda ke aiki kamar nau'in santsin lakabin dogaro da bayanai.
Wanne rarrabawa aka fi amfani da shi don samfurin hadawa coefficient lambda?
Mixup da CutMix yawanci suna zana λ daga rarrabawar Beta, wanda ke sarrafa yadda aka gauraya misalan biyu.