git clone git@github.com:hiwonjoon/atari-irl.git --recursive
conda env create -f environment.yml
conda activate atari-irl
pip install -r requirements.txt
Open rllab/sampler/stateful_pool.py and fix the typo.
from joblib.pool import MemmapingPool -> from joblib.pool import MemmappingPool
Open rllab/setup.py and change the line if package.startswith('rllab')], to if package.startswith('rllab') or package.startswith('sandbox')],
cd rllab
pip install -e .
cd baselines
check whether branch is correct. commit hash must be 24fe3d6. If not, please run git checkout 24fe3d6. Then,
pip install -e .
cd inverse_rl
create setup.py and put the code below.
# setup.py
from setuptools import setup,find_packages
setup(
name='inverse_rl',
packages=[package for package in find_packages()
if package.startswith('inverse_rl')],
version='0.1.0',
)
pip install -e .
- Train an Expert
python -m scripts.train_expert --env PongNoFrameskip-v4 --expert_path ./experts/pong --no-one_hot_code --num_envs 8 --n_cpu 8
- Generate Trajectory
python -m scripts.generate_trajectories --env PongNoFrameskip-v4 --expert_path ./experts/pong/checkpoints/update-100 --n_cpu 1 --num_envs 1 --no-render --num_trajectories 8
-
If you want to check the generate trajectories quality, run
python -m scripts.trajectory_to_gif --trajectories_file ./experts/pong/checkpoints/update-100/trajectories.pkl
- Train Auto-encoder
python -m scripts.train_ae --env PongNoFrameskip-v4 --num_envs 8 --encoder_type pixel_class --log_path ./autoencoder/pong
- Run AIRL algorithm
-
With trained AE
python -m scripts.train_airl --env PongNoFrameskip-v4 --n_cpu 1 --num_envs 1 --trajectories_file ./experts/pong/checkpoints/update-100/trajectories.pkl --no-state_only --encoder ./autoencoder/pong/vae_100.pkl --log_path ./airl/pong --reward_type mlp -
With raw ob
python -m scripts.train_airl --env PongNoFrameskip-v4 --n_cpu 1 --num_envs 1 --trajectories_file ./experts/pong/checkpoints/update-100/trajectories.pkl --no-state_only --log_path ./airl/pong