Why running train with --gpu 1.0 is not faster than without ?

thtrieu / darkflow

Translate darknet to tensorflow. Load trained weights, retrain/fine-tune using tensorflow, export constant graph def to mobile devices

GNU General Public License v3.0

6.13k stars 2.09k forks source link

Why running train with --gpu 1.0 is not faster than without ? #1189

Open ArnaudNorel95 opened 3 years ago

ArnaudNorel95 commented 3 years ago

Hello

I think I have correctly installed gpu's driver, CUDA and CUDNN according to this screenshot : cudacudnn

I tried training darkflow without GPU in a first time and it works. I can notice that GPU is indeed not running with nvidia-smi

Secondly I tried training darkflow with GPU (--gpu 1.0) and I notice that it is absolutely not faster. What is the interest ? Did I do something wrong ? With nvidia-smi I can notice that GPU is processing :

thank you for reading and if you could help me :)

andrewchang0910 commented 3 years ago

You probably not using your display card GPU, you are just use the GPU on motherboard. So you can not efficiently enhance the GPU usage. You should check your CUDA or CUDNN setting whether correct!