So the key issue with The Hummingbird Tutorial is that it's written for ML-Agents.1.0. There are some key changes that have to be made.
Really helpful in migrating the code to ML-Agents 2+
https://www.immersivelimit.com/tutorial ... -for-unity
https://www.immersivelimit.com/tutorial ... -ml-agents
In particular, the conversion of Heuristic and OnActionReceived functions which transitioned to ActionBuffers in HummingbirdAgent.cs.
I also had to upgrade the trainer_config.yaml file to the newest format for v2.0. I found on the Unity forum that a python script to automatcially migrates previous configs is provided:
https://forum.unity.com/threads/trainer ... id.912671/
These are specific to my setup on my home workstation/drives:
- Here's what I ran with my mlagents_env virtual environment active. I'm not sure if this works properly independent of venv.
python -m mlagents.trainers.upgrade_config /Volumes/T3/Unity/ml-agents/config/hummingbird/trainer_config.yaml /Volumes/T3/Unity/ml-agents/config/hummingbird/hummingbird_new.yaml
As opposed to the tutorial that uses Anaconda on a PC for venv, I'm using venv and running my training from inside the cloned ml-agents directory on T3.
mlagents-learn config/hummingbird/hummingbird_new.yaml --run-id=hummingbird
After those changes, ALL GOOD!