Netwọk squeeze-na-Excitation
Ihe mgbochi Squeeze-and-Excitation (SE) na-eme ka netwọkụ na-eme mgbanwe mụta oke ọwa njirimara ọ bụla, na-emegharị ha dabere na ọnọdụ ụwa.
Nchịkọta
This cheap attention-like mechanism won the 2017 ImageNet competition and became a standard CNN building block.
Ime miri emi
Ndị Hu, Shen na Sun webatara na 2017, ngọngọ SE na-agbakwunye anya ọwa doro anya na CNN. Ọ na-arụ ọrụ na nzọụkwụ abụọ. 'Mpịakọta' ahụ na-eji nkezi ọdọ mmiri zuru ụwa ọnụ daa maapụ atụmatụ ọ bụla (ịdị elu x obosara) n'ime otu ọnụọgụ, na-ewepụta otu nkọwa n'otu ọwa na-achịkọta ọrụ ya zuru ụwa ọnụ. The 'akpali akpali' na-azụ na vector site na abụọ obere njikọ zuru ezu n'ígwé na a bottleneck (a ReLU wee a sigmoid) na-emepụta a kwa ọwa arọ n'etiti 0 na 1. The arọ na-amụba mbụ atụmatụ map, amplifying bara uru ọwa na damping adịghị mkpa. SENet meriri ihe ịma aka nhazi ọkwa ILSVRC 2017, belata njehie top-5 ruo ihe dịka 2.25%. Ihe mgbochi a na-agbakwunye naanị nkeji ole na ole ma gbakọọ, yana oghere n'ime ResNet, Inception, ma ọ bụ MobileNet nwere obere mgbanwe.
Nghọta nka nka
Mpịakọta ahụ na-emepụta vector z-ogologo C ebe z_c bụ nkezi oghere nke ọwa c. Ngụkọta mkpali s = sigmoid (W2 * ReLU(W1 * z)), ebe W1 na-ebelata nha site na nbelata nha r (na-adịkarị 16) na W2 na-eweghachi ya, na-edobe ọnụ ahịa agbakwunyere obere. Nsonaazụ bụ maapụ atụmatụ ndenye nwere ọwa nke nwere amamihe site na s. Ọ bụ ụdị ịgbapụ onwe ya: netwọk na-ekpebi, site na ọnụ ọgụgụ zuru ụwa ọnụ, nke ọwa dị mkpa maka ntinye a kapịrị ọnụ.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke netwọkụ squeeze-and-Excitation
SE blocks na-ebi n'ime ụlọ arụ ọrụ nke ọma: EfficientNet na MobileNetV3 tinyere ha n'ime ụlọ ha. Echiche a na-amịpụta ezinụlọ nke modul nlebara anya, CBAM na-agbakwụnye nlebara anya gbasara ohere, ECA-Net na-eji mgbanwe 1D dị ọnụ ala dochie ọkpọ ahụ, na aghụghọ ndị a dị arọ na-apụta ugbu a na nchọpụta, nkewa, na ọbụna ụfọdụ ngwakọ transformer. Na-atụ anya nlebara anya ọwa ka ọ ga-abụ ihe nleba anya ziri ezi dị ọnụ ala ebe ọ bụla ọgba aghara dị.
Mmejuputa n'ezie n'ụwa
SENet meriri ihe ịma aka nhazi nke ImageNet ILSVRC 2017 site n'ịgbakwunye SE blocks na ọkpụkpụ azụ ResNeXt
EfficientNet na MobileNetV3 tinyere modul SE na ngọngọ ọ bụla iji kwalite izi ezi na ngwaọrụ mkpanaka
Ihe nchọpụta ihe na ụdị nkewa na-etinye ihe mgbochi SE iji mesie ọwa njirimara ozi ike
ECA-Net na CBAM na-agbatị echiche SE site na iji nhazigharị ọwa dị ọnụ ala ma ọ bụ nke mara mma
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
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
Ntuziaka na-esote
Netwọk Capsule
Ajụjụ a na-ajụkarị
What is Squeeze-and-Excitation Networks?
Ihe mgbochi Squeeze-and-Excitation (SE) na-eme ka netwọkụ na-eme mgbanwe mụta oke ọwa njirimara ọ bụla, na-emegharị ha dabere na ọnọdụ ụwa. Usoro nlebara anya dị ọnụ ala a meriri asọmpi ImageNet 2017 wee bụrụ ngọngọ ụlọ CNN ọkọlọtọ.
Kedu ihe nzọụkwụ 'squeeze' na ngọngọ SE na-eme?
Mpịakọta ahụ na-emetụta nchikota zuru ụwa ọnụ, na-akụda maapụ atụmatụ ọ bụla n'ime otu nkọwa n'otu ọwa.
Gịnị na-arụpụta nzọụkwụ 'excitation'?
Mmasị ahụ na-agafe vector a kpọgidere na-agafe n'ígwé abụọ FC na-ejedebe na sigmoid, na-enye nha nha ọwa ọ bụla.
Kedu otu esi etinye oke mkpali na maapụ atụmatụ izizi?
A na-amụba maapụ atụmatụ ọwa ọ bụla site n'ịdị arọ ọ mụtara, na-amụba ma ọ bụ na-egbochi ọwa ahụ.
Kedu ihe kpatara mkpali ahụ ji eji obere mbelata (mgbe 16)?
Mgbochi karama ahụ na-eweghachi akụkụ ọwa ahụ, yabụ ngọngọ SE na-efu naanị pasentị ole na ole.
Kedu asọmpi SENet meriri na 2017?
SENet meriri ihe ịma aka nhazi onyonyo nke ILSVRC 2017, na-akpali njehie top-5 ruo ihe dịka 2.25%.