Uhlaka lwe-Savant 0.2.7 Python lukhishiwe, okwenza kube lula ukusebenzisa i-NVIDIA DeepStream ukuxazulula izinkinga ezihlobene nokufunda komshini. Uhlaka lunakekela konke ukuphakamisa okusindayo nge-GStreamer noma i-FFmpeg, okukuvumela ukuthi ugxile ekwakhiweni kwamapayipi okukhiphayo alungiselelwe usebenzisa i-syntax ye-declarative (YAML) kanye nemisebenzi ye-Python. I-Savant ikuvumela ukuthi udale amapayipi asebenza ngokulinganayo kuma-accelerator esikhungweni sedatha (NVIDIA Turing, Ampere, Hopper) nakumadivayisi asemaphethelweni (NVIDIA Jetson NX, AGX Xavier, Orin NX, AGX Orin, New Nano). Nge-Savant, ungakwazi ukucubungula kalula ukusakazwa kwamavidiyo amaningi ngesikhathi esisodwa futhi udale ngokushesha amapayipi okuhlaziya amavidiyo alungele ukukhiqiza usebenzisa i-NVIDIA TensorRT. Ikhodi yephrojekthi isatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0.
I-Savant 0.2.7 inguqulo yakamuva yokukhishwa kwesici egatsheni le-0.2.X. Ukukhishwa okuzayo egatsheni le-0.2.X kuzofaka kuphela ukulungiswa kweziphazamisi. Ukuthuthukiswa kwezici ezintsha kuzokwenziwa egatsheni le-0.3.X, ngokusekelwe ku-DeepStream 6.4. Leli gatsha ngeke lisekele umndeni wakwaJetson Xavier wamadivayisi njengoba i-NVIDIA ingawasekeli ku-DS 6.4.
Okuqanjiwe okuyinhloko:
- Izimo ezintsha zokusetshenziswa:
- Isibonelo sokusebenza ngemodeli yokuthola esekelwe ku-transformer ye-RT-DETR;
- I-CUDA post-processing nge-CuPy ye-YOLOV8-Seg;
- Isibonelo sokuhlanganiswa kwe-PyTorch CUDA epayipini le-Savant;
- Ukuboniswa kokusebenza ngezinto eziqondiwe.
- Izici ezintsha:
- Ukuhlanganiswa ne-Prometheus. Ipayipi lingathumela ama-metrics okusebenzisa ku-Prometheus nase-Grafana ukuze kuqashwe ukusebenza nokulandelela. Onjiniyela bangamemezela amamethrikhi angokwezifiso athunyelwa kanye namamethrikhi esistimu.
- I-adaptha ye-Buffer - Isebenzisa isigcinalwazi esiqhubekayo kudiski sokuhambisa idatha phakathi kwama-adaptha namamojula. Ngosizo lwayo, ungathuthukisa amapayipi alayishwe kakhulu adla izinsiza ngokungalindelekile futhi amelane nokuqhuma kwethrafikhi. I-adaptha ithumela ingxenye yayo nedatha yosayizi ku-Prometheus.
- Imodi yokuhlanganisa imodeli. Amamojula manje angakwazi ukuhlanganisa amamodeli awo ku-TensorRT ngaphandle kokusebenzisa ipayipi.
- Isibambi somcimbi wokuvala shaqa we-PyFunc. Le API entsha ivumela ukuvalwa kwamapayipi ukuthi kuphathwe kahle, kukhululwe izinsiza futhi kwazise amasistimu ezinkampani zangaphandle ukuthi ukuvala shaqa kwenzeke.
- Ukuhlunga kozimele kokokufaka nokukhiphayo. Ngokuzenzakalelayo, ipayipi lamukela bonke ozimele abaqukethe idatha yevidiyo. Ngokuhlunga okokufaka nokukhiphayo, abathuthukisi bangakwazi ukuhlunga idatha ukuze bavimbele ukucutshungulwa.
- Ukucutshungulwa kwangemuva kwemodeli ku-GPU. Ngesici esisha, abathuthukisi bangakwazi ukufinyelela amamodeli okukhipha ama-tensor ngokuqondile kumemori ye-GPU ngaphandle kokuwalayisha kumemori ye-CPU futhi bawacubungule basebenzisa i-CuPy, i-TorchVision noma i-OpenCV CUDA.
- Imisebenzi yokumelela inkumbulo ye-GPU. Kulokhu kukhishwa, sinikeze imisebenzi yokuguqula izigcinalwazi zememori phakathi kwama-OpenCV GpuMat, PyTorch GPU tensor, nama-CuPy tensor.
- I-API yokufinyelela izibalo zokusetshenziswa kolayini bamapayipi. I-Savant ikuvumela ukuthi ungeze imigqa phakathi kwe-PyFuncs ukuze usebenzise ukucubungula okufanayo kanye nokucubungula ibhafa. I-API eyengeziwe inikeza onjiniyela ukufinyelela kolayini abafakwe epayipini futhi ibavumela ukuthi babuze ngokusetshenziswa kwabo.
Ekukhishweni okulandelayo (0.3.7) kuhlelwe ukuthuthela ku-DeepStream 6.4 ngaphandle kokwandisa ukusebenza. Umqondo uwukuthola ukukhululwa okuhambisana ngokugcwele ne-0.2.7, kodwa ngokusekelwe ku-DeepStream 6.4 kanye nobuchwepheshe obuthuthukisiwe, kodwa ngaphandle kokuphula ukuhambisana ezingeni le-API.
Source: opennet.ru