Question Why ML-Agents don't use all my system resources?

hello first of all, english is not my native language

i use ml-agents in my AI-RobotBattle game project

when i trained my AI, cpu usage in taskmanager too low
and framerate also too low compared to my pc spec

usually cpu usage when i training is about 10~20%

i hope ml-agents is use all my system resource propely

this trouble was happened both editor and build

and this is my pc spec

CPU : AMD Ryzen Threadripper 1950X 16Core
Ram : 64gb
VGA : RTX3090

 

Maxing out ml-agents CPU usage is hard because there exist several bottlenecks that are not CPU related, like exchanging data between Unity and Python. We are working on tools to help maximize usage but we do not have anything to show yet. My recommendation would be to increase the number of agents in the environment and also increase the number of environments used for training. This should increase the data throughput and increase the frequency of updates in the trainer.

 

thank you for answer!!

 

I have an environment that provides so many steps that the python seems to be severely bottlenecked by only being able to use a single CPU thread. Are there any options for increasing the amount of threads it can utilize for training?

 

Same here, I have one CPU maxed out while others hardly do any work.

 

I found the proper solution in this link:
ML-Agents: Multiple environments (num-envs) - Unity Forum
They said that's because "Pytorch is only configured to use 4 Threads at once" in ml-agents.

 

Well, I have not seen any way I can make mlagents use more threads. At least for the most important process that also takes the vast bulk of the memory and seems to manage everything.

I do use CUDA/GPU since it has performed better or almost identically without making my system hard to use while training. It does not utilize it fully either though. Either way, I would be interested in finding a way to make that main process use more cores, if that is even possible at the moment.

I am not remotely step/experience constrained which has usually been the argument for not having to bother with increasing threads for python/mlagents. I can get 100k expereiences in a minute though so that is not a bottleneck and I could greatly increase the scale of that.