I-Microsoft ivule umthombo womtapo wokusesha we-vector osetshenziswa ku-Bing

Microsoft eshicilelwe amakhodi omthombo womtapo wokufunda womshini SPTAG (Space Partition Tree And Graph) ngokusetshenziswa kwe-algorithm elinganiselwe ukusesha komakhelwane oseduze. Umtapowolwazi ithuthukisiwe ophikweni locwaningo lwe-Microsoft Research kanye nesikhungo sokuthuthukisa ubuchwepheshe bokusesha (Isikhungo Sezobuchwepheshe Se-Microsoft). Empeleni, i-SPTAG isetshenziswa injini yokusesha ye-Bing ukuze kutholwe imiphumela efaneleke kakhulu ngokusekelwe kumongo wemibuzo yosesho. Ikhodi ibhalwe ku-C++ futhi isatshalaliswa ngu ngaphansi kwelayisense ye-MIT. I-Build ye-Linux ne-Windows isekelwe. Kukhona ukubophezela kolimi lwePython.

Naphezu kweqiniso lokuthi umqondo wokusebenzisa i-vector storage ezinjinini zokusesha usunesikhathi eside uzulazula, empeleni, ukuqaliswa kwazo kuphazanyiswa ukushuba okuphezulu kwensiza yokusebenza ngama-vector kanye nemikhawulo yokuqina. Ukuhlanganisa izindlela zokufunda zomshini ezijulile nezilinganiso zokusesha zomakhelwane eziseduze kwenze kwaba nokwenzeka ukuletha ukusebenza nokuqina kwamasistimu e-vector ezingeni elamukelekayo ezinjinini ezinkulu zokusesha. Isibonelo, ku-Bing, kunkomba yevekhtha engaphezu kwezigidigidi eziyi-150, isikhathi sokulanda imiphumela efaneleke kakhulu siphakathi kuka-8 ms.

Umtapo wolwazi uhlanganisa amathuluzi okwakha inkomba nokuhlela ukusesha kwe-vector, kanye nesethi yamathuluzi okugcina uhlelo lokusesha olusabalalisiwe lwe-inthanethi oluhlanganisa amaqoqo amakhulu kakhulu ama-vector. Kunikelwe amamojula alandelayo: umakhi wenkomba wokukhomba, isisheli sokusesha sisebenzisa inkomba esabalaliswe kwiqoqo lamanodi amaningana, iseva yezibambi ezisebenzayo kumanodi, I-Aggregator yokuhlanganisa amaseva amaningana abe yinye, kanye neklayenti lokuthumela imibuzo. Ukufakwa kwama-vector amasha kunkomba kanye nokususwa kwama-vector endizeni kuyasekelwa.

Umtapo wolwazi usikisela ukuthi idatha ecutshungulwayo futhi yethulwa eqoqweni ifomethwe ngendlela yama-vector ahlobene angaqhathaniseka ngokusekelwe I-Euclidean (L2) noma cosine amabanga Umbuzo wosesho ubuyisela ama-vectors ibanga lawo phakathi kwawo nevekhtha yokuqala lilincane. I-SPTAG ihlinzeka ngezindlela ezimbili zokuhlela isikhala se-vector: I-SPTAG-KDT (isihlahla se-K-dimensional (kd-isihlahla) futhi igrafu yendawo ehlobene) kanye ne-SPTAG-BKT (k-isho isihlahla (k-usho umuthi kanye negrafu yendawo ehlobene). Indlela yokuqala idinga izinsiza ezincane lapho usebenza nenkomba, futhi eyesibili ibonisa ukunemba okuphezulu kwemiphumela yosesho yamaqoqo amakhulu kakhulu ama-vectors.

Ngesikhathi esifanayo, ukusesha i-vector akugcini embhalweni futhi kungasetshenziswa olwazini lwe-multimedia nezithombe, kanye namasistimu okukhiqiza izincomo ngokuzenzakalelayo. Isibonelo, enye yama-prototypes asekelwe ohlakeni lwe-PyTorch yasebenzisa isistimu ye-vector yokusesha ngokusekelwe ekufananeni kwezinto ezisezithombeni, ezakhiwe kusetshenziswa idatha evela kumaqoqo ereferensi amaningana anezithombe zezilwane, amakati nezinja, ezaguqulwa zaba amasethi ama-vector. . Uma isithombe esingenayo sitholwa ukuze siseshwe, siguqulwa kusetshenziswa imodeli yokufunda yomshini sibe ivekhtha, ngokusekelwe lapho amavekhtha afanayo kakhulu akhethwa kunkomba kusetshenziswa i-algorithm ye-SPTAG futhi izithombe ezihambisanayo zibuyiselwa njengomphumela.

Source: opennet.ru