Skip to content

installation.md

Latest commit

 

History

History
98 lines (94 loc) · 3.27 KB

installation.md

File metadata and controls

98 lines (94 loc) · 3.27 KB

Installation Instructions

Docker (Recommended)

  • clone this repo and then run the following commands in terminal from the root directory of this repo
  • create a docker image from the Dockerfile 
    CONTAINER_TAG="marl-jax-container"
docker build -t ${CONTAINER_TAG} -f Dockerfile .
  • running the docker image for a specific algorithm on meltingpot 
    ACTORS="2"
AVAIL_GPUS="0"
EXP_NAME="test"
ENV_NAME="meltingpot"
MAP_NAME="prisoners_dilemma_in_the_matrix__repeated"
ALGO="IMPALA"
STEPS="10000"
SEED="0"
docker run --gpus all --name ${EXP_NAME} -v $(pwd):/root/code ${CONTAINER_TAG} \
 train.py --async_distributed \
 --num_actors ${ACTORS} --available_gpus ${AVAIL_GPUS} \
 --algo_name ${ALGO} --env_name ${ENV_NAME} --map_name ${MAP_NAME} \
--num_steps ${STEPS} --seed ${SEED}

Local

  • create new conda environment

    conda create -n jax_rl python=3.9
  • [OPTIONAL] Install Cuda and CuDNN in conda if not available in system 
    • conda install cudnn # This automatically installs cuda too
    • conda install -c nvidia cuda-nvcc=11.3 # match the version of cuda installed in the above step

  • install requirements

    pip install -r requirements.txt

  • install lab2d

    • install using pip (link)

      pip install https://github.com/deepmind/lab2d/releases/download/release_candidate_2022-03-24/dmlab2d-1.0-cp39-cp39-manylinux_2_31_x86_64.whl
    • If the above fails, compile and install
      • Installing pre-requisites 
         conda install -c conda-forge gcc=12.2 gxx=12.2 bazel=5.2
      • Compiling lab2d 
         git clone https://github.com/deepmind/lab2d.git
cd lab2d
bazel build -c opt --define=lua=5_2 //dmlab2d:dmlab2d_wheel
        • Installing the compiled wheel file
         pip install bazel-bin/dmlab2d/dmlab2d-1.0-cp39-cp39-manylinux_._.._x86_64.whl

  • install meltingpot from git link

     git clone -b main https://github.com/deepmind/meltingpot
cd meltingpot
git reset --hard 9d3c74e68f9b506571706dd8be89e0809a8b4744
curl -L https://storage.googleapis.com/dm-meltingpot/meltingpot-assets-2.1.0.tar.gz \
   | tar -xz --directory=meltingpot
pip install -e .

  • install acme from git link

    git clone https://github.com/deepmind/acme
cd acme
git reset --hard 4525ade7015c46f33556e18d76a8d542b916f264
pip install -e .

  • Additional step is using conda environment

    # conda activate ENVNAME
# cd $CONDA_PREFIX
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
mkdir -p $CONDA_PREFIX/etc/conda/deactivate.d
echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/' > $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
echo 'unset LD_LIBRARY_PATH' > $CONDA_PREFIX/etc/conda/deactivate.d/env_vars.sh

  • Run IMPALA on prisoners_dilemma_in_the_matrix__repeated

     CUDA_VISIBLE_DEVICES=-1 python train.py --async_distributed --available_gpus 0 --num_actors 2 \
--algo_name IMPALA --env_name meltingpot --map_name prisoners_dilemma_in_the_matrix__repeated