Giters

keras-rl / keras-rl

Deep Reinforcement Learning for Keras.

Home Page:http://keras-rl.readthedocs.io/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Training performance is quite slow

Cpt-Falcon opened this issue · comments

commented

I was trying to run an something similar to the atari program but with real training. Training is very slow with an rtx 2070.

Keras RL is not using the nvidia optimized CUDNN libraries. Online i was seeing that the atari training example can take up to 40 hours on a tesla K40. That seems really high to me.

Is it expected for training to take that long, are there any optimizations that can accelerate things?

commented

You have to install tensorflow-gpu and change keras to tensorflow.keras
Hope this helps

commented

You have to install tensorflow-gpu and change keras to tensorflow.keras
Hope this helps

Thanks for the reply. Using tensorflow-gpu with an RTX 2070 cuts performance in half. I might not be using the cuDNN kernel and instead using the generic kernel, but still it shouldn't be that slow right?

commented

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.