PPO playing BipedalWalker-v3 from https://github.com/sgoodfriend/rl-algo-impls/tree/2067e21d62fff5db60168687e7d9e89019a8bfc0
4b89739
while getopts a:e:j:n:s:i: flag | |
do | |
case "${flag}" in | |
a) algo=${OPTARG};; | |
e) env=${OPTARG};; | |
j) n_jobs=${OPTARG};; | |
n) study_name=${OPTARG};; | |
s) seeds=${OPTARG};; | |
i) increment=${OPTARG};; | |
esac | |
done | |
TZ="America/Los_Angeles" | |
NOW=$(date +"%Y-%m-%dT%H:%M:%S") | |
study_name="${study_name:-$algo-$env-$NOW}" | |
STORAGE_PATH="sqlite:///runs/tuning.db" | |
increment="${increment:-100}" | |
mkdir -p runs | |
optuna create-study --study-name $study_name --storage $STORAGE_PATH --direction maximize --skip-if-exists | |
optimize () { | |
for ((j=$increment;j<=n_jobs*100+$increment;j+=100)); do | |
seed=() | |
for ((s=0;s<seeds;s++)); do | |
seed+="$((j+s*100/seeds)) " | |
done | |
echo python optimize.py --algo $algo --env $env --seed $seed --load-study --study-name $study_name --storage-path $STORAGE_PATH | |
done | |
} | |
optimize | xargs -I CMD -P $n_jobs bash -c CMD |