Darknet process not using all available GPU memory

[vladfulgeanu]

System Information Darknet: #80d9bec Compiled with CUDA 9.0, cuDNN, OpenCV 3.3.1 and openMP

Detailed description When processing a video for object detection, the darknet process only uses ~1181MiB out of the total 4029MiB available from the GPU.

The command issued is: ./darknet detector demo cfg/coco.data cfg/yolo.cfg weights/yolo.weights <video>.mp4

This can be seen using the nvidia-smi command:

Also the video runs at a very low fps i.e ~9 fps.

Is there a way to set how much GPU memory is used? Also is there any way to improve the fps other than decreasing the size in the cfg/yolo.cfg file?

[baristahell]

Well, isn't it normal? Considering it's :

• the loaded weights
• the activations for the current image/batch you can't really use an arbitrary amount of memory.

But the 9 fps thing is weird, maybe an issue with your video stream, i manage to run detection on 1920*1080 images in 40ms.

[vladfulgeanu]

I understand that using smaller weights or decreasing the input image size will help increase the fps, but are there any other things that can be improved or that I am missing? How does the activations for the current image/batch affect the fps, or what can it be wrong with the video stream?