Closed mnsmuts closed 5 years ago
Same error with offline training using the basic model and a demo file.
Hi @mnsmuts - thanks for the bug report. I think I see the problem, I'll work on a fix today.
Fixed in https://github.com/Unity-Technologies/ml-agents/pull/2498 - during a refactor, the step initialization got moved to RLTrainer
instead of Trainer
which explains why it didn't affect PPO training.
We'll work on some more automated testing to make sure there's better coverage on the BC Trainer code.
Many thanks
We're going to be doing a bugfix release on Monday with the fix for this.
Version 0.9.3 was released yesterday, which contains the fix for this. I'll close this issue now.
This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.
New installation of 0.9.2 with new conda py env. -using the UnitySDK and running the 3DBall -player brain on one of the agents Learning brain on the rest, academy with two brains (player and learner) Nothing unusual with the setup or command (mlagents-learn config\online_bc_config.yaml --train --slow)
Obviously I changed the brain to imitate line in the default onlineBC yaml file to 3DBallPlayer
Windows 10 ML-Agents 0.9.2 3DBall
Any ideas?
Just tried 0.9.1 with the hotfix and can get offline training working, then switched back to 0.9.2 and tried the same (with the correct env) and received the attrib error again.