jetson tx2 3fps why?

[hexiangquan]

i have test 416 ssd i can get 15fps. but use yolov3 just get 3 fps

[springkim]

yolov3 is based on darknet53. This model is not very fast on embedded system. Actually, i am also working on jetson tx2. So i will make custom model like tiny-darknet(darknet-reference). You need to change darknet53(yolov3.cfg) for fast detection.

[hexiangquan]

@springkim i will try use tensorRT

[RushNuts]

@springkim hi there! how much fps you have on tx2? That's yolo version and height&width you use?

[TheExDeus]

Has anyone really managed to port Yolov2 or v3 to tensorRT? Custom layers make it a pain.

[hexiangquan]

i have done it 10fps

[springkim]

Hi @hexiangquan Did you use YOLOv3 as TensorRT?

[hexiangquan]

yes

发自我的 iPhone

在 2018年4月3日，下午9:46，KimBomm notifications@github.com 写道：

Hi @hexiangquan Did you use YOLOv3 as TensorRT?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[TheExDeus]

Could you share some info how you did it?

[bufistov]

Hi @hexiangquan,

Could you please share you code?

[ahunde]

I have got roughly 14 fps with tiny YOLO v3, that is not satisfactory either.

[bufistov]

My main findings were:

• Ensure that max frequency is set for GPU: $ ~/jetson_clocks.sh --show should indicate that current GPU frequency is equal to max GPU frequency

• Use 'NCHW' order of input images, convolutions and poolings

[LiangHao92]

@hexiangquan Can you tell me the memory usage of yolo v3 on tx2?

[sharoseali]

@hexiangquan please reply .. many people are looking on you.... please

[ccosc]

I only get 1fps,that sucks.