koobas.hobune.stream
ML-Agent imitation training get slower and slower as it progresses
I have a very simple set up. I have a rectangular area (random width/length on each agent reset), a target that the agent needs to reach, the agent itself, and 2 brains. The agent can rotate left/right, can move forward/backward, and can jump (not used yet). One brain is the teacher and the other is the student as per the docs.
After starting up the python command, then playing the unity scene, everything is fine; completely responsive and running at ~120 fps. The moment that it hits 1000 steps and the first log entry in the python command is shown, Unity slows way down. My fps drops in half, my agent moves slower and it gets very sluggish. Over time, it keeps getting worse and worse to the point where my agent is moving about 1/10th the speed and my fps is down to the teens.
My computer's CPU and memory usage stay around the same even as Unity slows down, so I don't think I'm running out of resources.
Has anyone else seen this happen?
