Open Cyclones-Y opened 8 months ago
Well, as we adopt distributed reinforcement learning, we start actors and learner in different dockers. In this way, the actors only use the CPUs while the learner adopts the GPU for training. Each actor is in its own docker and equipped with its IP address for data transportation and model synchronization. In this way, many ip addresses are needed if you want to run multiple actors for higher efficiency. This is also due to that the gamecore we use is of low efficiency so we have to use multiple actors.
Well, as we adopt distributed reinforcement learning, we start actors and learner in different dockers. In this way, the actors only use the CPUs while the learner adopts the GPU for training. Each actor is in its own docker and equipped with its IP address for data transportation and model synchronization. In this way, many ip addresses are needed if you want to run multiple actors for higher efficiency. This is also due to that the gamecore we use is of low efficiency so we have to use multiple actors.
Thanks a lot for your answer, that means I only need to open different docker containers on one machine to deploy your project right?
Yes, but you have to use docker network to assign ip for each docker.
Forgive my ignorance. I didn't quite understand the configuration of the program