0

I am currently attempting to run an object detector called YOLO (darkflow).

I have installed CUDA 8.0, CudNN 5.1, tensorflow 1.0 and tensorflow-gpu (both installed via pip).

I'm testing YOLO by using the following command line:

sudo python3 ./flow --model cfg/yolo.cfg --load yolo.weights --demo dji_0004.MP4 --gpu 1.0

The problem is: even using --gpu 1.0 it runs at ~0.8 FPS.

While running it, I used nvidia-smi to monitor my GPU and this is what I get:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 375.82                 Driver Version: 375.82                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 650     Off  | 0000:01:00.0     N/A |                  N/A |
| 16%   40C    P0    N/A /  N/A |    444MiB /   980MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0                  Not Supported                                         |
+-----------------------------------------------------------------------------+

Apparently tensorflow isn't using running it on my GPU.

GTX 650 has 3.0 compute capability according to NVIDIA, so using tensorflow-gpu should be no problem.

talonmies
  • 70,661
  • 34
  • 192
  • 269
Caio Suzuki
  • 1
  • 1
  • It looks to me like its using your GPU. 444MB of your GPU memory is used, and there is a process running on it. – Robert Crovella Jan 14 '18 at 05:48
  • Did you install tensorflow-gpu version? check it using this https://stackoverflow.com/a/38580201/5330223 – Sumsuddin Shojib Jan 14 '18 at 05:54
  • 1
    Orthogonal to your question: why `sudo`? – Siyuan Ren Jan 14 '18 at 12:50
  • @RobertCrovella even when I stop running yolo (and before I ran it), my GPU memory used is still around 440MB. – Caio Suzuki Jan 14 '18 at 17:28
  • @Ultraviolet I installed tensorflow-gpu 1.0, and I'm not sure what I should be looking at in that link you sent, but reading it I was able to check that the only device that tensorflow recognizes is my CPU. – Caio Suzuki Jan 14 '18 at 17:30
  • @SiyuanRen I'm working in this with a classmate, and he tested it before me, so he sent me the exact command line he was using, I just copied it, not really sure why he used sudo – Caio Suzuki Jan 14 '18 at 17:36
  • If the code snippet does not list your GPU then `tensorflow` didn't detect your GPU. Try to fix that first, YOLO comes later. – Sumsuddin Shojib Jan 14 '18 at 17:59
  • @Ultraviolet I've been looking for a solution for this problem and I can't seem to find anything that works, do you know any method I can use to make tensorflow recognize my GPU? – Caio Suzuki Jan 16 '18 at 16:45
  • As you have an older GPU I am not sure this doc. will work but I think this will help https://www.nvidia.com/en-us/data-center/gpu-accelerated-applications/tensorflow/ – Sumsuddin Shojib Jan 16 '18 at 16:49
  • make sure cuda with appropriate version of cudnn are working then just `pip install tensorflow-gpu` – Sumsuddin Shojib Jan 16 '18 at 16:50

0 Answers0