Closed fanyuzeng closed 5 years ago
One of the reasons may be this program use only CPU (because of the original paper's setting).
And how many threads to use depends on your CPU computation power.
In my case, training nav_maze_static_01
took me more than 2 days on i7-3770 to converge.
You mean setting more thread can make convergence faster within CPU computation power? Are there parameters I can adjust in Deepmind lab to speed up the training process, like fast forward of TV?
I read A3C paper "Asynchronous methods for deep reinforcement learning" again. In general, setting more threads can speed up the training.
Are there parameters I can adjust in Deepmind lab to speed up the training process, like fast forward of TV?
I'm not familiar with setting DeelmindLab setting. And
In general, setting more threads can speed up the training.
is true when your CPU can afford to those many threads.
Ok, thank you so much ; )
My cpu can afford more than 30 threads.
Except stairway_to_melon, training other mazes takes long times much more than 24 hours with the config.THREAD equal to 8. I wonder whether there are some parameters of Deepmind lab I can adjust to speed up the training process. In addition, If I set config.THREAD=50, Can it speed up the process? or will it affect the episode-reward curve? Many thanks to you.