Microsoft e bula laeborari ea patlo ea vector e sebelisoang ho Bing

Microsoft e phatlalalitsoeng Likhoutu tsa mohloli oa laeborari ea ho ithuta mochini SPTAG (Space Partition Tree And Graph) ka ts'ebetsong ea algorithm e hakanyetsoang batla baahelani ba haufi. Laebrari ntshetswa pele karolong ea lipatlisiso ea Lipatlisiso tsa Microsoft le setsi sa nts'etsopele ea theknoloji ea patlo (Microsoft Search Technology Center). Ha e le hantle, SPTAG e sebelisoa ke enjine ea ho batla ea Bing ho fumana liphetho tse nepahetseng ho latela maemo a lipatlisiso. Khoutu e ngotsoe ka C ++ le ajoa ke tlas'a laesense ea MIT. Haha bakeng sa Linux le Windows e tšehetsoa. Ho na le tlamo bakeng sa puo ea Python.

Leha taba ea hore mohopolo oa ho sebelisa polokelo ea li-vector lienjineng tsa ho batla o ntse o phaphamala ka nako e telele, ts'ebetsong, ts'ebetsong ea bona e sitisoa ke matla a phahameng a ts'ebetso e nang le li-vector le mefokolo ea scalability. Ho kopanya mekhoa e tebileng ea ho ithuta mochine le hoo e batlang e le mekhoa ea ho batla moahelani e haufi ho entse hore ho khonehe ho tlisa ts'ebetso le scalability ea lisebelisoa tsa vector boemong bo amohelehang bakeng sa lienjineri tse kholo tsa ho batla. Mohlala, ho Bing, bakeng sa index ea vector ea li-vector tse fetang limilione tse likete tse 150, nako ea ho lata liphetho tse bohlokoa haholo e ka tlase ho 8 ms.

Laebrari e kenyelletsa lisebelisoa tsa ho haha ​​​​index le ho hlophisa lipatlisiso tsa li-vector, hammoho le lisebelisoa tsa ho boloka mokhoa oa ho batla oa inthaneteng o koahelang pokello e kholo haholo ea li-vector. E filwe dimmojule tse latelang: index builder for indexing, searcher for searched using the index distributed in the cluster of multiple nodes, server for running handlers on nodes, Aggregator for compunting's multiple server into one, le client bakeng sa ho romela lipotso. Ho kenyeletsoa ha li-vector tse ncha ho index le ho hlakoloa ha li-vector ho fofa hoa tšehetsoa.

Laeborari e fana ka maikutlo a hore data e sebetsitsoeng le ho hlahisoa pokellong e entsoe ka mokhoa oa li-vector tse amanang tse ka bapisoang ho ipapisitsoe le Euclidean (L2) kapa cosine bohole Potso ea patlo e khutlisa li-vector tseo sebaka se pakeng tsa tsona le vector ea mantlha e leng nyane. SPTAG e fana ka mekhoa e 'meli ea ho hlophisa sebaka sa vector: SPTAG-KDT (sefate sa K-dimensional)kd-sefate) le graph ea boahelani e amanang) le SPTAG-BKT (k-bolela sefate (k-e bolela sefate le kerafo ea boahelani e amanang). Mokhoa oa pele o hloka lisebelisoa tse fokolang ha o sebetsa le index, 'me oa bobeli o bontša ho nepahala ho holimo ha liphetho tsa lipatlisiso bakeng sa pokello e kholo haholo ea li-vectors.

Ka nako e ts'oanang, patlo ea vector ha e felle feela mongolong mme e ka sebelisoa ho litaba tsa multimedia le litšoantšo, hammoho le lits'ebetsong tsa ho iketsetsa likhothaletso. Mohlala, e 'ngoe ea li-prototypes tse ipapisitseng le moralo oa PyTorch e kentse ts'ebetso ea vector ea ho batla e ipapisitseng le ho ts'oana ha lintho litšoantšong, e hahiloeng ho sebelisoa data ho tsoa likolekeng tse' maloa tsa litšupiso tse nang le litšoantšo tsa liphoofolo, likatse le lintja, tse ileng tsa fetoloa lihlopha tsa li-vector. . Ha setšoantšo se kenang se amoheloa bakeng sa lipatlisiso, se fetoloa ho sebelisoa mokhoa oa ho ithuta oa mochine hore e be vector, ho latela hore na li-vector tse tšoanang haholo li khethoa ho index ho sebelisoa algorithm ea SPTAG 'me litšoantšo tse amanang le tsona li khutlisetsoa.

Source: opennet.ru

Eketsa ka tlhaloso