Mixup ak CutMix
Mixup ak CutMix ay pexe yu ñuy yokk ay done lañu, ñuy sos misaali tàggat yu bees ci boole ñaari nataal ak seeni etiket.
Résumé
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.
Plongeur bu xóot
Jaxas (Zhang ak ñeneen, 2017) dafay forme misaal bu bees bu melni x̃ = λ·x_a + (1-λ)·x_b ak etiketu ỹ jaxase ak benn λ, fu λ jëlee ci benn séddale Beta. Loolu dafay ñaax model bi mu doxalee ci anam wu jaar yoon ci diggante misaal yi, di yombal pexe yi ci dogal yi ak gëna suqali etalonnage bi. CutMix (Yun et al., 2019) lu moy loolu dafay dagg gox bu am kaare ci nataal B ba noppi teg ko ci nataal A; diisaayu etiket bi ñu ngi koy tëral ci pixel yi nataal bu nekk di joxe. Ndax CutMix dafay tëye gox yi jëm ci nataal yi (te du njaxasu xeetu nit), dafay tëye structure spatial bu am njariñ ci di forse model bi mu topp mbir yu bari ak pàrti. Ñari pexe yooyu dañuy nekk ay regularisatër yu am doole, yokk njubte ci ImageNet-scale benchmarks, gëna yokk robusteness ci ger ak dugal ay noon.
Gis-gis xarala
Ñari pexe yooyu yépp dañuy soppi mébet biy ñàkk, duñu soppi dugal bi kese. Etiket bi dafay nekk lu yomb, lu jaxaso, kon ñàkka entropi croix mooy λ-pondéré boole ñaari xeetu - ci dëgg-dëgg xeetu etiket bu nooy buñ takk ci njaxasu pixel. Ci CutMix, λ mingi méngoo ak paccu pixel yiñ soppiwul, ñu xayma ko ci yaatuwaayu boyet biñ dagg xaaj ko ak yaatuwaayu nataal bi yépp, loolu mooy tax proporsioŋu etiket bi méngoo ak bariwaayu nataal bu nekk.
njeextalu pexe
Njëgg ak budget
Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
dogal yu gëna leer
Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
Xool kalite
Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Ëlëgu yokk Mixup ak CutMix
Augmentation bu sukkandiko ci njaxas mooy leegi standard ci rëset yu dëgër yiy xaaj nataal yi, te dafay jàppale pipeline yu bees yi ngir transformatëri gis-gis, te dañuy faral di soxla regularisation bu diis. Gëstu dafay wéy ci anam yu xam-xam (lu melni, def ay dagg ci gox yu am xibaar), jaxase token-niveau ngir transformateur, ak yokk ci audio, bind, ak done 3D. Xaarandil pexe jaxase ngir des levier bu yomb ngir yokk njub, etalonnage, ak dëgër ndax architecture yi dañuy gëna xiif ci done.
Doxal ci àdduna dëgg
Taggat ImageNet classifiers ak CutMix ngir yokk top-1 njubte ak gëna baaxal barabu mbir yi.
Jëfandikoo Mixup ngir gëna suqali etalonnage model bi suko defee wóolu yiñ séentu gëna méngoo ak dëgg gi.
Transformatëri gis-gis yu bari (lu melni, DeiT) ak Mixup ak CutMix boole ngir tàggat ci done yu néew.
Yokkateg doole ci ger nataal yi ak duggal yi nekk ci bitti séddale ci sistemu xool yu am solo ci kaaraange.
Risk yi ak balustrade yi
Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Roadmap ngir samp gi
Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
Benchmark ci biir sargal ak done yu dëggu.
Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
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
Gis bi ci topp
Yokkateg waxtu test
Laaj yi ñuy faral di laaj
What is Mixup and CutMix Augmentation?
Mixup ak CutMix ay pexe yu ñuy yokk ay done lañu, ñuy sos misaali tàggat yu bees ci boole ñaari nataal ak seeni etiket. Mixup dafay interpoler nataal yépp ak etiket yi, CutMix dafay jël benn patch rectangle ci benn nataal ci beneen nataal ba noppi jaxase etiket yi ci barabu patch bi - ñoom ñaar dañuy wàññi overfitting ba noppi gëna dëgër.
naka la ko Mixup di defaree misaalu tàggat yaram bu bees?
Njaxas dafay forme x̃ = λx_a + (1-λ)x_b ba noppi jaxase etiket yi ak benn λ biñ jëlee ci séddale Beta.
Lan mooy wuutale CutMix ak Mixup?
CutMix dafay kopie gox bu rectangle ci benn nataal dem ci beneen, denc ëmbiit bu méngoo ci gox bi moo gën ñu boole ñaari nataal.
Ci CutMix, naka lañuy xamee diisaayu njaxas (lambda) ngir etiket yi?
Proportion etiket bi ci CutMix mingi aju ci yaatuwaayu boîte biñ paste ci nataal bi yépp, suko defee etiket yi di méngoo ak pixel yiñ mëna gis.
Lu weesu nataalu dugal bi, lan leneen la Mixup ak CutMix soppali?
Ñari xeeti xam-xam yooyu dañuy jaxase etiket yi itam, ñu defar ay target yu woyof yuy nuru xeetu etiketu nooy bu aju ci done.
Ban séddale lañuy faral di jëfandikoo ngir jël misaalu koeffisien lambda biy jaxase?
Mixup ak CutMix dañuy jël λ ci distribution Beta, mooy saytu ni ñaari misaal yi di jaxasoo.