Training too slow problem

[mamimaka]

Hello @huy-ha , thank you for your open-source and giving such clear guidance.

[mamimaka]

But when I try to train a decentralized multi-arm motion planner from scratch in our lab's server computer with

mkdir runs
python main.py --config configs/default.json --tasks_path tasks/ --expert_waypoints expert/ --num_processes 16 --name multiarm_motion_planner

I found that I could only use two threads to train, no matter how I changed the parameter 'num_processes'. And this situation made me have to wait for about 700s to accomplish a training step.

To solve this problem, I try to read the codes and have two results.
First, if I added the initialization code of CPU's number in the function initialize_ray(), which is defined in the utils.py, my cpu can be used more. But the training times of every step are not be short.
Second, I set num_processes in the script utils's function setup directly, also no help.

How can I fully use my CPUs to train this motion planner, this problem has troubled me for 3 weeks, I need your help.