Samfurin Yankewa
Samfurin datse yana raguwa cibiyar sadarwa ta jijiyoyi ta hanyar cire ma'auni ko gabaɗayan tsarin da ke ba da gudummawa kaɗan ga fitarwa.
Dubawa
It cuts size, memory, and compute cost while aiming to keep accuracy nearly intact.
Zurfafa nutsewa
Horarwar hanyoyin sadarwa na jijiyoyi galibi ana daidaita su: yawancin haɗin gwiwa suna ɗaukar ƙananan ma'auni waɗanda ke da kyar ke shafar tsinkaya. Yanke yana ganowa kuma yana cire waɗannan, yana barin mafi ƙarancin ƙima. Tsirewar da ba a tsara ba yana fitar da ma'aunin nauyi ɗaya, yana samar da ƙananan matrices waɗanda za'a iya matsawa sosai amma suna buƙatar kayan aiki na musamman ko ɗakunan karatu don a zahiri sauri. Tsare-tsare yana kawar da duka raka'a - neurons, kawunan hankali, tashoshi, ko yadudduka - yana ba da ƙaramin ƙima mai yawa wanda ke gudana cikin sauri akan kayan masarufi na yau da kullun. Girke-girke na yau da kullun shine madauki na maimaitawa: jirgin ƙasa, datse mafi ƙarancin mahimman sigogi ta wasu ma'auni (sau da yawa girman nauyi), sa'an nan kuma daidaitawa don dawo da daidaiton da aka rasa, maimaita har sai girman ko maƙasudin saurin ya cika. Yanke nau'i-nau'i ta dabi'a tare da ƙididdigewa da distillation a cikin bututun turawa.
Fahimtar Fasaha
Ƙimar mahimmanci ta yanke shawarar abin da za a yanke. Ma'auni mafi sauƙi shine girma - ƙananan ma'auni cikakke ana ɗauka ba su da amfani. Ƙarin ingantattun hanyoyin ƙididdige tasirin kowane nauyi akan asarar ta amfani da gradients ko tsari na biyu (na tushen Hessian), kamar yadda yake cikin mafi kyawun salon salon Likitan Kwakwalwa. Hasashen tikitin Lottery ya lura cewa cibiyoyin sadarwa masu yawa sun ƙunshi ƙananan ƙananan hanyoyin sadarwa waɗanda, horarwa daga farkon farawa, za su iya dacewa da cikakken samfurin - yana nuna yawancin hanyar sadarwa ba ta da yawa daga farko.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar Tsige Model
Ana ƙara yin datsewa ga manyan nau'ikan harshe, inda hanyoyin da aka tsara suke cire kawunan hankali, ƙwayoyin jijiya, har ma da yadudduka don dacewa da ƙira akan ƙananan GPUs da na'urori masu gefe. Hardware da kernels waɗanda ke amfani da ɓarna (kamar NVIDIA's 2: 4 tsararriyar sparsity) suna girma, suna yin dasawa mara tsari cikin sauri. Yi tsammanin za a haɗa tsintsin kai akai-akai tare da ƙididdigewa da distillation a zaman wani ɓangare na bututun matsawa mai sarrafa kansa waɗanda ke yin ƙayyadaddun latency, kuzari, da kasafin kuɗi na ƙwaƙwalwar ajiya.
Aiwatar da Gaskiyar Duniya
Matsa babban samfurin harshe don aiki akan GPU ɗin mabukaci ɗaya maimakon tarin uwar garke.
Slimming samfurin hangen nesa don ya dace a cikin ƙwaƙwalwar wayar hannu ko kyamarar da aka saka.
Cire jagororin hankali daga na'urar Transformer tare da ɗigon ma'aunin inganci.
Rage kuzarin ƙididdigewa da latency don manyan ayyukan zirga-zirga don rage farashin girgije.
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
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 Model Pruning 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
Jagora na gaba
AI Model Kulawa
Tambayoyin da ake yawan yi
What is Model Pruning?
Samfurin datse yana raguwa cibiyar sadarwa ta jijiyoyi ta hanyar cire ma'auni ko gabaɗayan tsarin da ke ba da gudummawa kaɗan ga fitarwa. Yana yanke girma, ƙwaƙwalwar ajiya, da ƙididdige farashi yayin da yake ƙoƙarin kiyaye daidaito kusan daidai.
Menene ainihin ra'ayin da ke bayan datsa samfurin?
Yanke yana yin amfani da juzu'i ta hanyar yanke ma'auni mai ƙarancin gudummawa ko raka'a don rage ƙirar yayin kiyaye mafi yawan daidaito.
Mene ne ya bambanta tsagewar da aka tsara daga tsarin da ba a tsara ba?
Tsarin datsa yana kawar da dukkan abubuwan da aka gyara, yana haifar da ƙananan ƙira masu yawa waɗanda ke gudana cikin sauri akan daidaitaccen kayan aiki, yayin da ba a tsara shi ba yana ba da ma'aunin nauyi ɗaya, yana haifar da ɓarna.
Me yasa pruning mara tsari sau da yawa ya kasa hanzarta samfurin akan kayan masarufi na yau da kullun?
Watsewar sifili ba sa fassara ta atomatik zuwa ƙananan ƙididdiga; na'ura mai yawa har yanzu tana sarrafa su sai dai idan an yi amfani da ƙwaya ko ƙararrawa na musamman.
Menene girke-girke na yau da kullun na pruning?
Juyawa mai jujjuyawa yana canzawa tsakanin cire sigogi marasa mahimmanci da daidaitawa don dawo da daidaito, a hankali ya kai girman ko maƙasudin saurin.
Menene Hasashen Tikitin Lottery ke iƙirari?
Hasashen ya lura cewa babban hanyar sadarwa mara kyau da aka zaɓa ('tikitin nasara'), wanda aka horar da shi daga farkon farkonsa, zai iya isa ga daidaiton cikakkiyar hanyar sadarwa.