Convolution yuñ mëna tàqale ci xóotaayu
Convolution yuñ mëna tàqale ci xóotaayu facteur convolution buñ miin ci ñaari jéego yu gëna xéewale, di dagg limu yokk yi ak parametre yi.
Résumé
They are the trick that lets neural networks run on phones and edge devices without melting the battery.
Plongeur bu xóot
Benn convolution buñ miin dafay jaxase ay leeral ci espace bi ak ci chaine yi ci benn operation bu dëgër, te loolu dafay seer lool. ab convolution buñu mëna tàqale ci xóotaayu xaaj lii ci ñaari pàcc. Bi njëkk mooy jéego bu xóot bi dafay jëfandikoo benn filtre bu ndaw ci chaine bu nekk, di jàpp motif spatial yi ci chaine bu nekk waaye du musa jaxase chaine yi. Ñaareel ba mooy jéego pointwise bi dafay jëfandikoo convolution 1x1 ngir boole chaine yi ci pixel bu nekk, jaxase xibaar chaine bi te doo xool dëkkandoo yi. Buñu dindi filtre spatial ci jaxase chaine, calcul total bi dafay wàññeeku bu baax, lu bari 8 ba 9 yoon ci filtre 3x3, ak perte bu ndaw ci njub. Faktorisasioŋ boobu mooy yax gi MobileNet ak Xception tëral.
Gis-gis xarala
Ngir kernel 3x3 buy màndargaal M chaine de entrée ba N sortie ci kaw benn carte, convolution standard dafay njëg lu tollu ci 9 yoon M yoon N yokk-yokk ci barab bu nekk. Version biñ mëna tàqale dafay njëg 9 yoon M ci wàllu xóotaayu yokk M yoon N ci wàllu 1x1. Rasio bi mingi tollu ci 1/N + 1/9, kon ngir N bu rëy, sakkanal bi dafay jege facteur spatial 1/9.
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 convolution yuñ mëna tàqale ci xóotaayu
Convolutions yuñ mëna tàqale ci xóotaayu desnañu nekk xeetu gis-gis bu am solo te dañuy gëna feeñ ci jëmmal CNN-transformatër yu melni MobileViT ak ConvNeXt blocks. Lu IA bi ci aparey bi di màgg, gaawaayu hardware bi dafay yokk ndimmbalu ops yu xóot yi. Xaarandi jëfandikoo bu wéy ci xoolum ci jamono dëgg, kaptër yuñ mëna sol, ak bépp jekkal bu latency, mémoire, ak budget energie yu sew, ñu koy faral di boole ak seetlu kantite ak architecture neuronal.
Doxal ci àdduna dëgg
MobileNet ak MobileNetV2 dañu leen di jëfandikoo ngir doxal tànneefi nataal yi ci telefon yu xarañ yi ak latency bu néew
Segmentasioŋ portrait ci jamono dëgg ak lëndëm ci aplikaasioŋu woote wideo dañu sukkandikoo ci yaxu ndigg yu woyof yuñ mëna tàqale
Gis mbir ci biir aparey ci kamera kaaraange ak drone, fu doole ak ordinatër néew
Xception daf leen di jëfandikoo ci eskaal ngir gëna dëgëral ImageNet bi ñuy saytu limu parametre yi
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 Depthwise Separable Convolutions 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
Konvolusioŋ yuñ mëna soppi
Laaj yi ñuy faral di laaj
What is Depthwise Separable Convolutions?
Convolution yuñ mëna tàqale ci xóotaayu facteur convolution buñ miin ci ñaari jéego yu gëna xéewale, di dagg limu yokk yi ak parametre yi. Mooy pexe biy tax reso neuronal yi mëna dox ci telefon yi ak ci aparey yi te batëri bi du seey.
Lan mooy ñaari jéego yi convolution buñ mëna tàqale ci xóotaayu xaaj convolution buñ miin?
Jéego bu xóot bi dafay segg chaine bu nekk ci boppam, ba noppi jéego 1x1 biy boole ay leeral ci chaine yi.
Ci jéego bu xóot bi, naka lañuy jappe chaine de input yi?
Convolution xóotaayu dafay jëfandikoo filtre spatial bu wuute ci chaine bu nekk te du jaxase chaine yi, moo tax mu yomb.
Lu tollu ci ñaata la convolution 3x3 buñu mëna tàqale ci xóotaayu suuf mëna wàññi calcul bi suñu ko méngale ak convolution 3x3 buñ miin bu am chaine de sortie yu bari?
Ngir kernel 3x3 ratio sakkanal bi mingi jegesi 1/9 sudee limu chaine yiy génn dafa bari, loolu mooy joxe lu tollu ci 8 ba 9 yoon liggéey yu néew.
Lan mooy wareefu convolution pointwise (1x1) ci jëmmal bii?
Convolution pointwise 1x1 dafay jaxase chaine yi ci pixel bu nekk, te jéego bu xóot bi moo ko moytu def.
Ban architecture bu siiw moo siiwal convolution yuñ mëna tàqale ci xóotaayu aparey mobile yi?
MobileNet dañu ko tabax ci kaw ay convolution yuñ mëna tàqale ci xóotaayu suuf, ngir mëna am inference bu baax ci telefon yi.