Our goal is to run jaxmarl_testdrive.py. This is a basic script that you should only be able to run if you have everything installed correctly, which should be the case by the end of this section. I've only tested this out on macOS with an M1 chip. I suggest using mamba instead of conda.

1. Make a conda environment with conda create -n siggame python=3.9
2. Activate the environment with conda activate siggame
3. Install jax. If you're on macOS, you can run this if you are using conda:

   conda install -c conda-forge jaxlib=0.4.19
   conda install -c conda-forge jax

   Or you can run this if you are using mamba (better):

   mamba install jaxlib=0.4.19
   mamba install jax

   Verify that it is installed correctly by running python -c 'import jax; print(jax.numpy.arange(10))' You may get some warnings but you should end up with a list of 0-9.
4. Install JaxMARL. This cannot be done with pip, since we need to run the algorithms (See JaxMARL installation instructions for why). Clone JaxMARL in a separate directory with:

   git clone https://github.com/FLAIROx/JaxMARL.git && cd JaxMARL
   pip install -e .

5. If you're on mac, you may still struggle to run jaxmarl_testdrive.py because you'll get a MuJoCo error. However, we don't really care about MuJoCo, so you can just comment out the following lines in the JaxMARL package (I hate how dirty this is):

   # In /JaxMARL/jaxmarl/environments/__init__.py comment out line 20
   ...
   from .overcooked import Overcooked, overcooked_layouts
   # from .mabrax import Ant, Humanoid, Hopper, Walker2d, HalfCheetah
   from .hanabi import HanabiGame
   ...

   # In /JaxMARL/jaxmarl/registration.py comment out lines 19-23
   ...
   HeuristicEnemySMAX,
   LearnedPolicyEnemySMAX,
   SwitchRiddle,
   # Ant,
   # Humanoid,
   # Hopper,
   # Walker2d,
   # HalfCheetah,
   InTheGrid,
   InTheGrid_2p,
   HanabiGame,
   ...

   This should work for JaxMARL v0.0.2. If you're using a different version, you may need to find the lines yourself.
6. We also need torch and torchvision, which can be installed with mamba:

   mamba install torchvision

   The conda analog would be something like (untested):

   conda install -c conda-forge torchvision

7. We also need jax-dataloader, which can be installed with pip: We no longer need jax-dataloader.

   pip install jax-dataloader

Finally, we should be able to run jaxmarl_testdrive.py with:

python jaxmarl_testdrive.py

Follow the Oscar Jax Instructions, which are copied below.

1. Run the following:

       module purge
       unset LD_LIBRARY_PATH
       module load cuda cudnn

2. In a separate directory, create a python virtual environment:

       python -m venv jax.venv

3. Activate it: source jax.venv/bin/activate
4. Load jax:

       pip install --upgrade pip
       pip install --upgrade "jax[cuda12_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

   Verify by running python -c 'from jax.lib import xla_bridge;print(xla_bridge.get_backend().platform)' You should see gpu.
5. Install JaxMARL in a separate directory:

   git clone https://github.com/FLAIROx/JaxMARL.git && cd JaxMARL
   pip install -e .

6. Install torchvision (this may take a while):

   pip install torchvision

You should be able to run jaxmarl_testdrive.py.

Check the readme.md in the base_experiment/ folder for more information on that experiment.