FLAIROx / JaxMARL

Multi-Agent Reinforcement Learning with JAX
Apache License 2.0
395 stars 69 forks source link

started a pr about rnn hidsize #56

Closed alexunderch closed 7 months ago

alexunderch commented 7 months ago
  1. Fixed the Dockerfile — i.e. added important env. variables and eased development interaction with git;
  2. Changed IPPO baseline for SMAX — i.e. made it independent of rnn hidden_size. Say if the fix actually works and then I can implement it everywhere in IPPO/MAPPO baselines (in configs there is GRU_HIDDEN_SIZE variable (name is arbitrary), with default value of 256). Default experiment is accessible via:
    python ippo_rnn_smax.py GRU_HIDDEN_SIZE=128
  3. now WANDB_ENTITY could be parsed directly from the docker environment
  4. A small fix in the walkthrough colab — i.e. a note about runtime reboot after the installation.

If I have to add tests or smth, ping me.

amacrutherford commented 7 months ago

looks great! Have run the RNN change and results look reasonable