InoteveraGaidhi rinotevera
Optimizer State Offloading kuCPU uye NVMe
Tekinoroji
Nhungamiro yehunyanzvi
Tarisa mberi uye Shumba Optimizers
Lookahead neShumba maviri emazuva ano twist pane neural-network optimization.
Teerera kugwaro iriAI-yakarondedzerwa · ~2 min
Pa peji ino2 min verenga
Pfupiso
Lookahead inoputira chero base base optimizer ine 'inononoka' uye 'inokurumidza' huremu hwekufambira mberi kwakadzikama, nepo Shumba (EvoLved Sign Momentum) yakawanikwa nekutsvaga chirongwa cheAI uye inogadziridza uremu uchishandisa chete chiratidzo chenguva yekumhanya - ichiita kuti ive ndangariro-mwenje uye kazhinji nekukurumidza kupfuura Adhamu.
Kudzika Kwakadzika
Lookahead, yakakurudzirwa naZhang, Hinton uye vaanoshanda navo muna 2019, anomhanyisa yakajairwa 'nekukurumidza' optimizer (saAdhama kana SGD) yematanho ek, yobva yakwenya seti yakaparadzana 'inononoka' uremu chidimbu chenzira yakananga kwakaguma uremu hwekukurumidza. Izvi zvinoderedza oscillations uye kuderedza kunzwa kune hyperparameters. Shumba, yakaburitswa na Google muna 2023, yakabuda muchirongwa chekutsvaga pamusoro pe optimizer algorithms. Iyo inoteedzera kukurumidza asi inoshandisa chiratidzo chekushanda kune yekuvandudza, saka yega yega parameter inofamba neyakagadziriswa nhanho saizi munzira yeakaunganidzwa gradient chiratidzo. Shumba inochengeta chete nhanho yenhanho (hafu yenyika yaAdhamu, inochengeta mbiri), inoshandisa kuora uremu hombe uye mwero wekudzidza mudiki, uye yakafananidza kana kurova Adamu pakuona kukuru uye mienzaniso yemitauro painodzidzira nekukurumidza uye isingadhuri.
Kuitwa muNyika yeChokwadi
Kuputira Adamu ne Lookahead kudzikamisa kudzidziswa kwevashanduri uye kuderedza hyperparameter tuning kuedza.
Kushandisa Shumba kudzidzisa mamodheru makuru ekuona (semuenzaniso, ViT) ine yakaderera optimizer ndangariro kupfuura Adamu.
Kufanodzidzira mamodheru emitauro neShumba kuti iwane kurongeka kwakaenzaniswa pamutengo wakaderera wekombuta.
Kubatanidza Lookahead neSGD mukusimbisa-yekudzidza vamiririri kupfava zvine ruzha zvigadziriso zvepolicy.
Ramangwana reKutarisa uye Shumba Optimizers
Shumba yakagamuchirwa mune akati wandei mahombe ekudzidzira kumhanya nekuti inocheka optimizer ndangariro uye inokwanisa kukurumidzira kusangana, uye kuwanikwa kwayo kunoratidza otomatiki 'AI-designing-AI' algorithm yekutsvaga seyo chaiyo sosi yezvinobudirira zvinowana. Tarisira mamwe ma-optimizers ekutsvaga, mahybrid schemes anosanganisa Lookahead-maitiro anononoka uremu ane masign-based updates, uye kuwedzera kufarira mundangariro-anoshanda optimizers semhando dzemhando dzinoramba dzichisimbisa GPU memory bhajeti.
Kuzvishandisa
Nei zvichikosha
- Mutengo uye bhajeti. Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
- Sarudzo dzakajeka. Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
- Kudzora kwemhando yepamusoro. Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.
Chenjera
- Kugadzirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba.
- Infrastructure uye mari yekugadzirisa inowanzotarisirwa pasi.
- Chengetedzo uye kucherechedzwa mapundu anogona kukura sezvo masisitimu anowedzera kuoma.
Matanho anotevera
- Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
- Benchmark pasi pechokwadi mutoro uye data mamiriro.
- Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
- Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.
Ramba Uchiongorora
Tsamba yemahara
Tora iyo yemazuva ese AI pfupiso
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Take the Lookahead and Lion Optimizers quiz
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Mibvunzo inowanzo bvunzwa
Chii chinonzi Lookahead uye Shumba Optimizers?
Lookahead neShumba maviri emazuva ano twist pane neural-network optimization. Lookahead inoputira chero base base optimizer ine 'inononoka' uye 'inokurumidza' huremu hwekufambira mberi kwakadzikama, nepo Shumba (EvoLved Sign Momentum) yakawanikwa nekutsvaga chirongwa cheAI uye inogadziridza uremu uchishandisa chete chiratidzo chenguva yekumhanya - ichiita kuti ive ndangariro-mwenje uye kazhinji nekukurumidza kupfuura Adhamu.
Ndeipi inotsanangura nhanho yekombuta muShumba optimizer uremu hwekuvandudza?
Shumba inoshandisa chiratidzo kune imwe nguva inopindirana, saka imwe neimwe parameter inofamba nenhanho yakafanana mugwara rechiratidzo che gradient.
Lookahead inogadzirisa sei kugadzirisa kwayo?
Lookahead inomhanyisa inner optimizer yema k nhanho, yobva yafamba zvishoma nezvishoma uremu chidimbu chenzira yakananga kune zvinokurumidza uremu uye kuseta zvakare.
Iyo Lion optimizer yakatanga kuwanikwa sei?
Shumba (EvoLved Sign Momentum) yakabuda kubva muchirongwa-chekutsvaga maitiro akachinja uye akaongorora zvirongwa zvemumiriri.
Kuenzaniswa naAdamu, ndeipi yakakosha ndangariro mukana weShumba?
Adamu anoteedzera nguva yekutanga neyechipiri; Shumba inongochengeta imwe chete yekumhanyisa buffer, ingangoita nepakati nepakati optimizer state memory.
Sei Shumba ichiwanzoda mwero wekudzidza mudiki pane waAdam?
Sezvo basa rechiratidzo rinogadzirisa hukuru hwekugadzirisa kwega kwega, danho rinoshanda rakakura, saka chiyero chiduku chekudzidza (uye kuora kwehuremu) kunoshandiswa.
Ramba uchidzidza
Mazano anoenderana
Mamwe madhairekitori akasarudzirwa nyaya iyi