{"id":113469,"date":"2024-02-08T22:02:49","date_gmt":"2024-02-08T20:02:49","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/vypusk-savant-0-2-7-frejmvorka-kompyuternogo-zreniya-i-glubokogo-obucheniya"},"modified":"2024-02-08T22:02:49","modified_gmt":"2024-02-08T20:02:49","slug":"vypusk-savant-0-2-7-frejmvorka-kompyuternogo-zreniya-i-glubokogo-obucheniya","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/vypusk-savant-0-2-7-frejmvorka-kompyuternogo-zreniya-i-glubokogo-obucheniya","title":{"rendered":"Shp\u00ebrndarja Savant 0.2.7, nj\u00eb korniz\u00eb p\u00ebr vizionin kompjuterik dhe m\u00ebsimin e thell\u00eb","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
\u00cbsht\u00eb publikuar versioni 0.2.7 i framewok-it Python Savant, i cili thjeshton p\u00ebrdorimin e NVIDIA DeepStream p\u00ebr t\u00eb zgjidhur detyrat q\u00eb lidhen me m\u00ebsimin e makinerive. Framework-u merr p\u00ebrsip\u00ebr t\u00eb gjith\u00eb pun\u00ebn e komplikuar me GStreamer ose FFmpeg, duke lejuar p\u00ebrq\u00ebndrimin n\u00eb nd\u00ebrtimin e pipeline-ve t\u00eb optimizuar duke p\u00ebrdorur sintaks\u00ebn deklaruese (YAML) dhe funksionet Python. Savant lejon krijimin e pipeline-ve q\u00eb punojn\u00eb nj\u00ebsoj si n\u00eb akselerator\u00ebt n\u00eb qendrat e t\u00eb dh\u00ebnave (NVIDIA Turing, Ampere, Hopper), ashtu edhe n\u00eb pajisjet edge (NVIDIA Jetson NX, AGX Xavier, Orin NX, AGX Orin, New Nano). Me ndihm\u00ebn e Savant, \u00ebsht\u00eb e leht\u00eb t\u00eb p\u00ebrpunosh disa rrjedha video n\u00eb t\u00eb nj\u00ebjt\u00ebn koh\u00eb, duke krijuar shpejt pipeline t\u00eb gatshme p\u00ebr aplikacione videoanalitik\u00eb q\u00eb p\u00ebrdorin NVIDIA TensorRT. Kodi i projektit shp\u00ebrndahet n\u00ebn licenc\u00ebn Apache 2.0. <\/p>\n
Savant 0.2.7 \u2014 \u00ebsht\u00eb versioni i fundit me ndryshime funksionaliteti n\u00eb deg\u00ebn 0.2.X. Publikimet e ardhshme n\u00eb deg\u00ebn 0.2.X do t\u00eb p\u00ebrfshijn\u00eb vet\u00ebm rregullime gabimesh. Zhvillimi i funksioneve t\u00eb reja do t\u00eb vazhdoj\u00eb n\u00eb deg\u00ebn 0.3.X, e cila \u00ebsht\u00eb e bazuar n\u00eb DeepStream 6.4. Kjo deg\u00eb nuk do t\u00eb mb\u00ebshtes\u00eb familjen e pajisjeve Jetson Xavier, pasi kompania NVIDIA nuk i mb\u00ebshtet ato n\u00eb DS 6.4. <\/p>\n
T\u00eb rejat kryesore: <\/p>\n
\n
Shembuj t\u00eb rinj t\u00eb p\u00ebrdorimit:\n
\n
Shembull i pun\u00ebs me modelin e detektimit t\u00eb bazuar n\u00eb transformatorin RT-DETR;\n
Post-procesimi CUDA me CuPy p\u00ebr YOLOV8-Seg;\n
Shembulli i integrimit t\u00eb PyTorch CUDA n\u00eb pipeline-in Savant;\n
Demostrimi i pun\u00ebs me objekte t\u00eb orientuara. <\/ul>\n
<\/center> <\/p>\n
Ve\u00e7orit\u00eb e reja:\n
\n
Integrimi me Prometheus. Pipeline-i mund t\u00eb eksportoje metrikat e ekzekutimit n\u00eb Prometheus dhe Grafana p\u00ebr monitorimin dhe gjurmimin e performanc\u00ebs. Zhvilluesit mund t\u00eb shpallin metrikat e tyre t\u00eb personalizuara, t\u00eb eksportuara p\u00ebrkrah metrikave sistemike.\n
Adaptori i memorie \u2014 implementon nj\u00eb burim konstant transaksional n\u00eb disk p\u00ebr t\u00eb dh\u00ebnat q\u00eb l\u00ebvizin midis adapter\u00ebve dhe moduleve. Me t\u00eb, mund t\u00eb zhvilloni konvujer\u00eb me ngarkesa t\u00eb larta, q\u00eb konsumojn\u00eb ndjesh\u00ebm burime dhe durojn\u00eb shp\u00ebrthime n\u00eb trafik. Adapteri eksporton n\u00eb Prometheus t\u00eb dh\u00ebnat e tij n\u00eb lidhje me element\u00ebt dhe sizet.\n
Reprimi i kompilimit t\u00eb modeleve. Tani modulet mund t\u00eb kompilojn\u00eb modelet e tyre n\u00eb TensorRT pa e ekzekutuar pipeline-in.\n
P\u00ebrpunuesi i ngjarjes shutdown n\u00eb PyFunc. Ky API i ri lejon menaxhimin e duhur t\u00eb operacioneve t\u00eb p\u00ebrfundimit t\u00eb pipeline-it, duke liruar burimet dhe duke njoftuar sistemet e jashtme p\u00ebr p\u00ebrfundimin e pun\u00ebs.\n
Filtrimi i \u043a\u0430\u0434rov\u00ebve n\u00eb hyrje dhe dalje. Me parazgjedhje, konvejeri pranon t\u00eb gjitha \u043a\u0430\u0434rov\u00ebt q\u00eb p\u00ebrmbajn\u00eb t\u00eb dh\u00ebna video. Me ndihm\u00ebn e filtrimit n\u00eb hyrje dhe dalje, zhvilluesit mund t\u00eb filtrojn\u00eb t\u00eb dh\u00ebnat p\u00ebr t\u00eb p\u00ebrjashtuar p\u00ebrpunimin e tyre.\n
Pas p\u00ebrpunimi i modelit n\u00eb GPU. Fal\u00eb funksionit t\u00eb ri, zhvilluesit mund t\u00eb ken\u00eb qasje n\u00eb tenzor\u00ebt e daljes s\u00eb modelit direkt nga memorja e GPU, pa i ngarkuar ato n\u00eb memorjen e CPU dhe ti p\u00ebrpunojn\u00eb ato me CuPy, TorchVision ose OpenCV CUDA.\n
Funksionet e paraqitjes s\u00eb memories GPU. N\u00eb k\u00ebt\u00eb l\u00ebshim, kemi ofruar funksione p\u00ebr t\u00eb transformuar bufer\u00ebt e memories mes OpenCV GpuMat, tenzor\u00ebve GPU PyTorch dhe tenzor\u00ebve CuPy.\n
API p\u00ebr qasje n\u00eb statistikat e p\u00ebrdorimit t\u00eb radh\u00ebve t\u00eb konvejerit. Savant lejon shtimin e radh\u00ebve midis PyFunc p\u00ebr t\u00eb realizuar p\u00ebrpunimin paralel dhe buferizimin e p\u00ebrpunimit. API e shtuar ofron zhvilluesve qasje n\u00eb radh\u00ebt e despletuara n\u00eb konvejer dhe lejon k\u00ebrkimin e p\u00ebrdorimit t\u00eb tyre. <\/ul>\n<\/ul>\n
N\u00eb versionin e ardhsh\u00ebm (0.3.7), \u00ebsht\u00eb planifikuar kalimi n\u00eb DeepStream 6.4 pa zgjerimin e funksionalitetit. Ideja \u00ebsht\u00eb t\u00eb kemi nj\u00eb publikim plot\u00ebsisht t\u00eb p\u00ebrputhsh\u00ebm me 0.2.7, por t\u00eb bazuar n\u00eb DeepStream 6.4 dhe teknologji t\u00eb p\u00ebrmir\u00ebsuar, por pa shkelur p\u00ebrputhshm\u00ebrin\u00eb n\u00eb nivelin API. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![\"Shp\u00ebrndarja](\"\/wp-content\/uploads\/2024\/02\/0c14b808a6767394352a4eafde317f67.jpg\")
<\/center> <\/p>\n