Skip to content


Repository files navigation

Catalyst.RL: A Distributed Framework for Reproducible RL Research

Paper & Framework


System requirements

sudo apt install -y redis
sudo apt install -y python3-dev zlib1g-dev libjpeg-dev \ 
    cmake swig python-pyglet python3-opengl libboost-all-dev \
    libsdl2-dev libosmesa6-dev patchelf ffmpeg xvfb

Python env setup

conda create -n rl python=3.6 anaconda
source activate rl
conda remove nb_conda_kernels -y
conda install -c conda-forge nb_conda_kernels -y
conda install notebook jupyter nb_conda -y
conda remove nbpresent -y

Python requirements

pip install gym['all']
pip install -r ./requirements.txt


Local run - LunarLander

# terminal 1 - db node
redis-server --port 12000

# terminal 2 and 3
export GPUS=""  # like GPUS="0" or GPUS="0,1" for multi-gpu training
export CONFIG=./gym_lunarlander/sac_d3pg.yml  # or "td3_qd4pg.yml", "qd4pg.yml" 

# terminal 2 - trainer node
CUDA_VISIBLE_DEVICES="$GPUS" catalyst-rl run-trainer --config="${CONFIG}"

# terminal 3 - samplers node
CUDA_VISIBLE_DEVICES="" catalyst-rl run-samplers --config="${CONFIG}"

# terminal 4 - progress visualization
CUDA_VISIBLE_DEVICE="" tensorboard --logdir=./logs

Benchmark - BipedalWalker

export GPUS=""  # like GPUS="0" or GPUS="0,1" for multi-gpu training
export EXP_DIR="gym_bipedalwalker_simple"  # or "gym_bipedalwalker_hardcore"
export CONFIG=./_base/_all.yml,./_base/_agents101.yml,./_base/_qd4pg.yml,./"${EXP_DIR}"/qd4pg.yml,./_base/_ddpg.yml
export LOGDIR=./logs/"${EXP_DIR}"/ddpg-qd4pg

    --redis-port 12100 \
    --config "$CONFIG" \
    --logdir "$LOGDIR" \
    --param-name "shared/n_step" \
    --param-values "1, 5" \
    --param-type "int" \
    --wait-time 10800 \  # 3 hours, use 43200 for 12 hours experiment 
    --n-trials 1  # number of trials per experiment  


Please cite the following paper if you feel this repository useful.

  title={Catalyst.RL: A Distributed Framework for Reproducible RL Research},
  author = {Kolesnikov, Sergey and Hrinchuk, Oleksii},
  journal={arXiv preprint arXiv:1903.00027},

Related Projects


For any question, please contact

Sergey Kolesnikov:
Oleksii Hrinchuk: