YoloV3-Tiny outperform YoloV3-Tiny-Xnor?

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.63k stars 7.95k forks source link

YoloV3-Tiny outperform YoloV3-Tiny-Xnor? #3095

Open rtrahms opened 5 years ago

rtrahms commented 5 years ago

Hi - Thanks for providing a Tiny-xnor cfg to evaluate! Much appreciated. We noticed that after training two networks with the exact same image dataset (using yolo-cpp-dll and a wrapper application), with CUDA 10 on an NVidia P2000 GPU (1024 CUDA cores), during testing we achieved 100-110 FPS on YoloV3-Tiny and only 90 FPS on YoloV3-Tiny-Xnor. We were surprised that xnor was slower! Is there a build step that we missed on using the XNOR network? The only change that we made was the addition/setting of the xnor flag in the cfg file.

Any help/guidance would be appreciated! Rob

AlexeyAB commented 5 years ago

@rtrahms Hi,

XNOR net yet in progress. At the moment I am targeting Larger networks, by using more modern GPUs. There are still bottlenecks, especially for small (Tiny) networks: https://github.com/AlexeyAB/darknet/issues/3054

Try to train https://github.com/AlexeyAB/darknet/files/2853459/yolov3-spp_xnor_obj.cfg.txt with width=416 height=416 at least 4000-5000 iterations and check FPS.

More: https://github.com/AlexeyAB/darknet/issues/2365#issuecomment-462923756