search

LAMDA-RL / ODIS

The implementation of ICLR-2023 paper "Discovering Generalizable Multi-agent Coordination Skills from Multi-task Offline Data".
Apache License 2.0
38 stars 5 forks source link

NaN during Cooperative Navigation Tasks #4

Open euikk opened 1 week ago

euikk commented 1 week ago

Hello, I have an NaN issue while running Cooperative Navigation Tasks

NaN occurs during train_vae. I just change following setting:

Is it just a setting issue? then please let me know the settings you used for the ICLR paper. (other baselines work well)

Thanks!

mansicer commented 5 days ago

I ran the experiment in my machine and didn't find the nan issue. Here is the command:

python src/main.py --mto --config=odis --env-config=cn_offline --task-config=cn-expert --t_max=30000 --entity_embed_dim=32 --attn_embed_dim=64 --skill_dim=4 --seed=1

Maybe you can check

  1. if you have updated to the latest code via git pull (I fixed some wrong datasets in recent few commits)
  2. whether the versions of torch, numpy, and others correspond to the requirements.
euikk commented 4 days ago

Thank you. Till now, it works well.

I have several questions.

  1. In your ICLR 2023 paper table 6, the hyper-parameter 'attention dimension' is 64. Does it mean that 'attn_embed_dim' is 64? (in your code, it is 8 originally)

  2. I think NaN is because of entitiy_embed_dim. Then for MPE, the settings below are what you use for paper?

Thanks again.