Ikibazo gihuza ibibazo · Hagati Urwego
Igihombo Cyibanze kubibazo bitaringanijwe
Kugenzura icyerekezo-gutakaza modulator, kutaringaniza ibyiciro, uruhare rwa alfa na gamma, nuburyo ingero zashyizwe hamwe ningero zikomeye zitandukanye.
Ikibazo 1 Bya 8
Ni ikihe kibazo cy'ibanze mu cyiciro kimwe cyo gutahura cyari igihombo cyibanze kugirango gikemuke?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Ni ikihe kibazo cy'ibanze mu cyiciro kimwe cyo gutahura cyari igihombo cyibanze kugirango gikemuke?
- Ubusumbane bukabije hagati yibintu bidasanzwe byimbere nibintu byoroshye byoroshye
- Buhoro buhoro umuvuduko wibice bibiri byerekana
- Kubura amakuru yanditseho amahugurwa
- Kurenza ibintu binini
- Ni ikihe kintu cyerekana ko igihombo cyibanze kigwira kuri cross-entropy?
- log (p_t)
- (1 - p_t) ^ gamma
- p_t ^ 2 gusa
- 1 / p_t
- Bigenda bite kubura igihombo mugihe kwibanda kuri gamma byashyizwe kuri 0?
- Ihinduka ikosa rya kare
- Irengagije ingero zose nziza
- Ihinduka cross-entropy isanzwe
- Itandukana
- Niki kongera gamma ikora kurugero rworoshye, rwashyizwe mubikorwa neza?
- Yongera umusanzu wabo
- Kureka umusanzu wabo udahindutse
- Kubikura kuri dataset
- Kurwanya umusanzu wabo wigihombo cyane
- Ni ubuhe bwubatsi bwa detector bwatangije igihombo cyibanze?
- RetinaNet
- YOLOv1
- Byihuta R-CNN
- AlexNet
- Ni uruhe ruhare ibipimo bya alfa bisanzwe bigira hamwe na gamma?
- Ishiraho igipimo cyo kwiga
- Iringaniza uburemere bugereranije nibyiza hamwe nibyiciro bibi
- Igenzura imiterere yishusho
- Irasobanura umubare wibisanduku
- For a hard example with p_t close to zero, what is the focal modulating factor (1 − p_t)^gamma likely to do?
- Drive the factor to zero
- Turn the example into a positive label
- Keep the factor relatively close to one so the example retains substantial loss weight
- Replace the cross-entropy term with alpha
- A dense detector has many easy background locations. What effect does focal loss aim to have on their combined training influence?
- Increase their count by duplicating them
- Remove all positive object examples
- Make every example receive an identical loss
- Reduce their dominance so training can focus more on sparse hard examples