How to improve the GPU-Util when running RL program with RLLab.

[morantumi]

Hi,
I have tried my best to archieve this goal but i failed. I found that theano＝0.9.0　pygpu=0.6.5 and theano=1.0.2 pygpu=0.7.5 can pass the test program and the result is showed as follow. I think it can prove that theano can work correctly.

However, when I run python program in rllab with the setting use_gpu=True, I still have the bad result like following:

I think the config of theano is ok because the test program is running correctly. And ~/.theanorc is setted with device=cuda0.
When I run program with setting use_gpu=False, everything is ok. So I think the problem is about the Theano with GPU. So I'm really confused and please tell me how to use thean with GPU to accelerate rllab.
Thank you very much.

[morantumi]

I solve the problem.
However, the GPU-Util is too low
I use the command watch -n 0.1 nvidia-smi and find the GPU-Util is about 3%~5%. And most of time is 0%. I don't know how to deal with it. could you give me some advice?

[qxtian]

can you tell me the solution for the problem- "you are tring to use the old GPU back-end'? thanks.