Microsoft inatsegula laibulale yosakira vekitala yomwe imagwiritsidwa ntchito mu Bing

Microsoft losindikizidwa makina ophunzirira library library code SPTAG (Space Partition Tree And Graph) ndikukhazikitsa ma algorithm oyerekeza kusaka koyandikana nawo pafupi. Library otukuka mu gawo lofufuza la Microsoft Research ndi malo otukula ukadaulo wosaka (Microsoft Search Technology Center). M'malo mwake, SPTAG imagwiritsidwa ntchito ndi injini yosakira ya Bing kuti idziwe zotsatira zoyenera kwambiri kutengera zomwe zafufuzidwa. Khodiyo imalembedwa mu C ++ ndi wogawidwa ndi pansi pa MIT layisensi. Kumanga kwa Linux ndi Windows kumathandizidwa. Pali chomangira chilankhulo cha Python.

Ngakhale lingaliro la kugwiritsa ntchito ma vector kusungirako mu injini zosaka lakhala likuyandama kwa nthawi yayitali, mukuchita, kukhazikitsidwa kwawo kumalepheretsedwa ndi kuchulukitsitsa kwazinthu zogwirira ntchito ndi ma vectors ndi malire a scalability. Kuphatikiza njira zophunzirira zamakina zakuya ndi njira zofufuzira zoyandikana nawo pafupi zapangitsa kuti zitheke kubweretsa magwiridwe antchito ndi ma scalability a makina a vector pamlingo wovomerezeka pamainjini akulu osakira. Mwachitsanzo, mu Bing, pa vekitala ya ma vector opitilira 150 biliyoni, nthawi yotengera zotsatira zoyenera kwambiri ili mkati mwa 8 ms.

Laibulaleyi imaphatikizapo zida zomangira chilozera ndi kukonza kusaka kwa ma vector, komanso zida zingapo zosungira makina osakira pa intaneti omwe amaphatikiza magulu akulu kwambiri a ma vector. Zoperekedwa magawo otsatirawa: omanga indexing for indexing, searcher for searching pogwiritsa ntchito index yogawidwa mgulu la node zingapo, seva yoyendetsa ma node, Aggregator yophatikiza ma seva angapo kukhala amodzi, ndi kasitomala potumiza mafunso. Kuphatikizika kwa ma vector atsopano mu index ndikuchotsa ma vector pa ntchentche kumathandizidwa.

Laibulale imatanthawuza kuti zomwe zasinthidwa ndikuperekedwa muzosonkhanitsa zimasinthidwa kukhala ma vector ogwirizana omwe angafanane ndi Euclidean (L2) kapena kodi mtunda Funso losaka limabweza ma vector omwe mtunda wake pakati pawo ndi vesi loyambirira ndi lochepa. SPTAG imapereka njira ziwiri zokonzera malo vekitala: SPTAG-KDT (K-dimensional mtengo (K-dimensional tree)kd - mtengo) ndi wachibale graph) ndi SPTAG-BKT (k-kutanthauza mtengo (k-amatanthauza mtengo ndi graph yoyandikana nayo). Njira yoyamba imafuna zinthu zochepa pogwira ntchito ndi ndondomekoyi, ndipo yachiwiri imasonyeza kulondola kwapamwamba kwa zotsatira zakusaka pazosonkhanitsa zazikulu kwambiri za ma vectors.

Nthawi yomweyo, kusaka vekitala sikungokhala pamawu ndipo kumatha kugwiritsidwa ntchito pazambiri zamawu ndi zithunzi, komanso pamakina opangira zopangira zokha. Mwachitsanzo, imodzi mwama prototypes ozikidwa pa PyTorch framework idakhazikitsa njira yofufuzira potengera kufanana kwa zinthu zomwe zili pazithunzi, zomwe zidapangidwa pogwiritsa ntchito zidziwitso zochokera m'magulu angapo okhala ndi zithunzi za nyama, amphaka ndi agalu, zomwe zidasinthidwa kukhala ma vector. . Chithunzi chomwe chikubwera chikalandiridwa kuti chifufuzidwe, chimasinthidwa pogwiritsa ntchito makina ophunzirira makina kukhala vekitala, kutengera zomwe ma vector ofanana kwambiri amasankhidwa kuchokera pamndandanda pogwiritsa ntchito algorithm ya SPTAG ndipo zithunzi zomwe zikugwirizana nazo zimabwezeretsedwa.

Source: opennet.ru

Kuwonjezera ndemanga