search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.72k stars 7.96k forks source link

0 FPS testing on video file running on K40 GPU #3991

Open ani16 opened 5 years ago

ani16 commented 5 years ago

Hello Alex.

Referring your mail below mentioned on thread 80 @Isabek You can achive about ~0.01 FPS per 1 GFlops-SP using yolo-voc.cfg network 416x416 on my fork.

The network is trained with VOC using yolo-voc.cfg with weight and height 416. But while testing on video file it shows ~0 FPS only. Why is it so when K40 has 4291–5040 SP GFLOPS means around 40-50 FPS it can provide.

Each frame takes 27.46 GFLOPS . Please tell me if I am going wrong somewhere.

Thanks in advance.

ani16 commented 5 years ago

Hi Alex

Please do reply on this too as I am facing problem with fps. Each frame takes ~65 GFLOPS with yolov3. I am not getting enough fps.

ani16 commented 5 years ago

One more thing, please tell me

@Isabek You can achive about ~0.01 FPS per 1 GFlops-SP using yolo-voc.cfg network 416x416 on my fork. What does it mean .How have you calculated?

AlexeyAB commented 5 years ago

when K40 has 4291–5040 SP GFLOPS means around 40-50 FPS it can provide.

Yes.

But while testing on video file it shows ~0 FPS only.

You didn't use GPU. Compile Darknet with CUDA.