Configure StarCraft II environment for DeepMind challenge on NYU Prince Cluster

README.md

Prerequisites

0. Load singularity module:

module load singularity/2.4.4

1. Download and unzip StarCraft II for Linux: here I take StarCraft II 3.16 for example, you could replace the link and zip file name according to SC2 Official Linux Downloads

   1. Download zip

   wget http://blzdistsc2-a.akamaihd.net/Linux/SC2.3.16.1.zip

   2. Unzip

   unzip SC2.3.16.1.zip

   3. Prompt the password iagreetotheeula
   4. Move StarCraft II files

   mv ./StarCraftII/ $SCRATCH/StarCraftII

2. Download and unzip StarCraft II mini maps

   1. Download mini game maps

   wget https://github.com/deepmind/pysc2/releases/download/v1.2/mini_games.zip

   2. Unzip

   unzip mini_games.zip

   3. Move mini game maps

   mv ./mini_games $SCRATCH/StarCraftII/Maps/

3. Start the Ubuntu container

singularity shell --nv /beegfs/work/public/singularity/ubuntu-17.10.img

4. Create and activate virtual environment

   1. Create virtual environment named py3tf

   virtualenv -p python3 $SCRATCH/py3tf

   2. Activate virtual environment py3tf

   source $SCRATCH/py3tf/bin/activate

5. Add python dependencies in virtual environment

   1. Install tensorflow-gpu

   pip install tensorflow-gpu

   2. Install pysc2

   pip install pysc2

   3. (Optional) Install other dependencies defined in requirements.txt

   pip install -r requirements.txt

6. Deactivate virtual environment, exit the container and return to login node

   1. Deactivate virtual environment

   deactivate

   2. Exit container

   exit

Batch mode(Preferred)

1. Download the model you wanna train, I take this one for example
   1. Download

   git clone https://github.com/simonmeister/pysc2-rl-agents

   2. Move

   mv pysc2-rl-agents $SCRATCH/

2. Download slurm scripts
   1. Download and move control code(you should modify the notification email address): modify code as you see fit(references1,2)

   wget https://gist.github.com/jimmyahacker/1bb8b99b58a9d2af1ca8582f9f8d14dc/raw/d5e5d2a71ebfaa3a18e938c6e58baf127ad57bca/job.s;mv job.s $SCRATCH/

   2. Download and move in-container code

   wget https://gist.github.com/jimmyahacker/1bb8b99b58a9d2af1ca8582f9f8d14dc/raw/d5e5d2a71ebfaa3a18e938c6e58baf127ad57bca/in_container.sh; mv in_container.sh $SCRATCH/

3. Run in batch mode

cd $SCRATCH;sbatch job.s

4. See job status

squeue -u $USER

Interactive mode(ONLY FOR DEBUGGING USE)

0. Download allocate_gpus.sh written below

   1. Download

   wget https://gist.github.com/jimmyahacker/1bb8b99b58a9d2af1ca8582f9f8d14dc/raw/d11fa4148bb506103b08a2b44fe06247fb1f472b/allocate_gpus.sh

   2. Move to home directory

   mv ./allocate_gpus.sh $SCRATCH/allocate_gpus.sh

1. Load packages and allocate GPUs: allocate 1 GPU(you could replace 1 with any number of GPUs you wanna allocate)

source $SCRATCH/allocate_gpus.sh 1

2. Start the Ubuntu container

singularity shell --nv /beegfs/work/public/singularity/ubuntu-17.10.img

3. Run virtual environment py3tf

source $SCRATCH/py3tf/bin/activate

4. Set StarCraftII environment variable

export SC2PATH="$SCRATCH/StarCraftII"

allocate_gpus.sh

#!/bin/bash
module purge
module load cuda/9.0.176 cudnn/9.0v7.0.5 singularity/2.4.4
export MY_LD_LIBRARY_PATH=$LD_LIBRARY_PATH
srun --gres=gpu:$1 --pty /bin/bash

in_container.sh

#!/bin/bash
cd $SCRATCH
source py3tf/bin/activate
cd pysc2-rl-agents
python run.py my_experiment --map MoveToBeacon --envs 1 --iters 1

job.s

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --gres=gpu:1
#SBATCH --time=10:00:00
#SBATCH --mem=3GB
#SBATCH --job-name=GPU-SC2-Demo
#SBATCH --mail-type=ALL
#SBATCH --mail-user=@nyu.edu
#SBATCH --output=slurm_%j.out
module purge
module load cuda/9.0.176 cudnn/9.0v7.0.5 singularity/2.4.4
export MY_LD_LIBRARY_PATH=$LD_LIBRARY_PATH
export SC2PATH=$SCRATCH/StarCraftII
singularity exec --nv /beegfs/work/public/singularity/ubuntu-17.10.img bash in_container.sh

Requirements

PySC2 Old version

used in here

absl-py==0.1.13
astor==0.6.2
bleach==1.5.0
enum34==1.1.6
future==0.16.0
futures==3.1.1
gast==0.2.0
google-apputils==0.4.2
grpcio==1.10.1
html5lib==0.9999999
Markdown==2.6.11
mock==2.0.0
numpy==1.14.2
pbr==4.0.1
portpicker==1.2.0
protobuf==3.5.2.post1
pygame==1.9.3
PySC2==1.1
python-dateutil==2.7.2
python-gflags==3.1.2
pytz==2018.3
s2clientprotocol==1.1
six==1.11.0
tensorboard==1.7.0
tensorflow-gpu==1.7.0
termcolor==1.1.0
websocket-client==0.47.0
Werkzeug==0.14.1