Reseaux de squeeze-et-excitation
Bloc Squeeze-and-Excitation (SE) dafay may reso convolutionnel mu jàng ba ñaata lay wara def ci chaine bu nekk, di leen kalibrewaat ci anam wi àdduna bi di doxee.
Résumé
This cheap attention-like mechanism won the 2017 ImageNet competition and became a standard CNN building block.
Plongeur bu xóot
Hu, Shen, ak Sun ñoo ko dugal ci 2017, blok SE dafay yokk chaine bu leer ci CNN. Dafay dox ci ñaari jéego. 'Squeeze' dafay jëfandikoo pooling moyenne global ngir dajale bépp kàrtu man-man (kawewaay x yaatuwaay) ci benn nimero, génne benn tegtal ci chaine bu nekk buy tënk aktivaasioŋ global bi. 'Excitation' bi dafay dundal vecteur bi jaar ci ñaari couche yu ndaw yu lëkkaloo bu baax ak benn bottleneck (ReLU ginaaw ga sigmoid) ngir génne poid ci chaine bu nekk diggante 0 ak 1. Poid yooyu dañuy yokk kàrtu màndarga yi njëkk, di yokk chaine yu am njariñ yi ak damping yi amul njariñ. SENet moo jël ndam li ci ILSVRC 2017, dagg njuumte top-5 ba lu tollu ci 2.25%. Blok bi du yokk ludul ay parametre yu néew ci teemeer boo jél ak xayma, ba noppi dugg ci ResNet, Inception, wala MobileNet ak coppite yu néew.
Gis-gis xarala
Squeeze bi dafay defar vecteur z bu am guddaay C, z_c mooy moyenne spatiale bu chaine c bi. Excitation dafay xayma s = sigmoid (W2 * ReLU (W1 * z)), fu W1 wàññi dimension ci ratio wàññi r (dafay faral di nekk 16) ak W2 defaraat ko, baña yokk njëg bi. Lépp luy génne mooy kàrtu man-mani duggal buñ scaleer chaine-wise ci s. Dafay nuru self-gating: reso bi mooy jël dogal, ci lim yi ci àdduna bi, ban chaine moo am solo ci input bi.
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 Reseau Squeeze-ak-Excitation
Blok SE yi ñu ngi dundu ci biir tabax yu am njariñ: EfficientNet ak MobileNetV3 dañu leen samp ci seeni blok tabax. Xalaat bi dafa jur famiy modulu bàyyi xel, CBAM yokk bàyyi xel ci barab, ECA-Net dafa wecci bottleneck bi ak benn convolution 1D bu yomb, te yii pexe recalibration yu woyof leegi dañu feeñ ci gis, segmentation, ba ci yenn hybrids vision-transformer. Xaarandil ni bàyyi xel ci chaine bi nekk levier bu yomb njëg fépp fu convolution yi wéy.
Doxal ci àdduna dëgg
SENet moo jël ndam li ci ImageNet ILSVRC 2017 ndax dafa yokk ay bloku SE ci yaxu ndigg bu ResNeXt
EfficientNet ak MobileNetV3 dañuy dugal ay modle SE ci bépp blok ngir gëna dëgëral aparey mobile yi
Detektëru mbir yi ak xeetu xaaj yi dañuy dugal ay blok SE ngir wane chaine yu am solo yi
ECA-Net ak CBAM dañu yokk xalaatu SE ci etalonnage chaine bu gëna yomb wala bu xam barab bi
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 Squeeze-and-Excitation Networks 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
Reseau capsule
Laaj yi ñuy faral di laaj
What is Squeeze-and-Excitation Networks?
Bloc Squeeze-and-Excitation (SE) dafay may reso convolutionnel mu jàng ba ñaata lay wara def ci chaine bu nekk, di leen kalibrewaat ci anam wi àdduna bi di doxee. Bii mekaniism bu yomb te nuru luñuy bàyyi xel moo jël ndam li ci ImageNet 2017 ba noppi nekk bloku tabax buñ miin ci CNN.
Lan la jéego 'squeeze' bi ci benn blok SE di def?
Squeeze bi dafay jëfandikoo pooling moyenne global, di dajale kàrtu man-man bu nekk ci benn tegtal ci chaine bu nekk.
Lan mooy jéego 'excitation' biy génne?
Excitation dafay jaar ci vecteur biñ tass ci ñaari couche FC yuy jeex ci sigmoïde, loolu mooy joxe poids scaling ci chaine bu nekk.
naka lañuy jëfandikoo poids excitation yi ci kàrtu màndarga yi njëkk?
Bépp kàrtu màndarga chaine dañu koy yokk ci diisaay bimu jàng, di yokk wala di wàññi chaine boobu.
Lan moo waral excitation bi di jëfandikoo ab bottleneck bu am ratio buy wàññi (dafay faral di nekk 16)?
Bottleneck bi dafay wàññeeku ba noppi defaraat dimension chaine bi, kon SE block bi dafay yokk tuuti ci teemeer boo jél.
Ban competition la SENet jëlee ci 2017?
SENet moo jël ndam li ci ILSVRC 2017 ci tànneefi xaaj nataal yi, puus top-5 njuumte ci lu tollu ci 2.25%.