RL Baselines3 Zoo is a training framework for Reinforcement Learning (RL), using Stable Baselines3.
It provides scripts for training, evaluating agents, tuning hyperparameters, plotting results and recording videos.
In addition, it includes a collection of tuned hyperparameters for common environments and RL algorithms, and agents trained with those settings.
We are looking for contributors to complete the collection!
Goals of this repository:
- Provide a simple interface to train and enjoy RL agents
- Benchmark the different Reinforcement Learning algorithms
- Provide tuned hyperparameters for each environment and RL algorithm
- Have fun with the trained agents!
This is the SB3 version of the original SB2 rl-zoo.
The hyperparameters for each environment are defined in hyperparameters/algo_name.yml
.
If the environment exists in this file, then you can train an agent using:
python train.py --algo algo_name --env env_id
For example (with tensorboard support):
python train.py --algo ppo --env CartPole-v1 --tensorboard-log /tmp/stable-baselines/
Evaluate the agent every 10000 steps using 10 episodes for evaluation (using only one evaluation env):
python train.py --algo sac --env HalfCheetahBulletEnv-v0 --eval-freq 10000 --eval-episodes 10 --n-eval-envs 1
Save a checkpoint of the agent every 100000 steps:
python train.py --algo td3 --env HalfCheetahBulletEnv-v0 --save-freq 100000
Continue training (here, load pretrained agent for Breakout and continue training for 5000 steps):
python train.py --algo a2c --env BreakoutNoFrameskip-v4 -i rl-trained-agents/a2c/BreakoutNoFrameskip-v4_1/BreakoutNoFrameskip-v4.zip -n 5000
When using off-policy algorithms, you can also save the replay buffer after training:
python train.py --algo sac --env Pendulum-v1 --save-replay-buffer
It will be automatically loaded if present when continuing training.
Plot scripts (to be documented, see "Results" sections in SB3 documentation):
scripts/all_plots.py
/scripts/plot_from_file.py
for plotting evaluationsscripts/plot_train.py
for plotting training reward/success
Examples (on the current collection)
Plot training success (y-axis) w.r.t. timesteps (x-axis) with a moving window of 500 episodes for all the Fetch
environment with HER
algorithm:
python scripts/plot_train.py -a her -e Fetch -y success -f rl-trained-agents/ -w 500 -x steps
Plot evaluation reward curve for TQC, SAC and TD3 on the HalfCheetah and Ant PyBullet environments:
python3 scripts/all_plots.py -a sac td3 tqc --env HalfCheetahBullet AntBullet -f rl-trained-agents/
The RL zoo integrates some of rliable library features. You can find a visual explanation of the tools used by rliable in this blog post.
First, you need to install rliable.
Note: Python 3.7+ is required in that case.
Then export your results to a file using the all_plots.py
script (see above):
python scripts/all_plots.py -a sac td3 tqc --env Half Ant -f logs/ -o logs/offpolicy
You can now use the plot_from_file.py
script with --rliable
, --versus
and --iqm
arguments:
python scripts/plot_from_file.py -i logs/offpolicy.pkl --skip-timesteps --rliable --versus -l SAC TD3 TQC
Note: you may need to edit plot_from_file.py
, in particular the env_key_to_env_id
dictionary
and the scripts/score_normalization.py
which stores min and max score for each environment.
Remark: plotting with the --rliable
option is usually slow as confidence interval need to be computed using bootstrap sampling.
The easiest way to add support for a custom environment is to edit utils/import_envs.py
and register your environment here. Then, you need to add a section for it in the hyperparameters file (hyperparams/algo.yml
).
Note: to download the repo with the trained agents, you must use git clone --recursive https://github.com/DLR-RM/rl-baselines3-zoo
in order to clone the submodule too.
If the trained agent exists, then you can see it in action using:
python enjoy.py --algo algo_name --env env_id
For example, enjoy A2C on Breakout during 5000 timesteps:
python enjoy.py --algo a2c --env BreakoutNoFrameskip-v4 --folder rl-trained-agents/ -n 5000
If you have trained an agent yourself, you need to do:
# exp-id 0 corresponds to the last experiment, otherwise, you can specify another ID
python enjoy.py --algo algo_name --env env_id -f logs/ --exp-id 0
To load the best model (when using evaluation environment):
python enjoy.py --algo algo_name --env env_id -f logs/ --exp-id 1 --load-best
To load a checkpoint (here the checkpoint name is rl_model_10000_steps.zip
):
python enjoy.py --algo algo_name --env env_id -f logs/ --exp-id 1 --load-checkpoint 10000
To load the latest checkpoint:
python enjoy.py --algo algo_name --env env_id -f logs/ --exp-id 1 --load-last-checkpoint
Upload model to hub (same syntax as for enjoy.py
):
python -m utils.push_to_hub --algo ppo --env CartPole-v1 -f logs/ -orga sb3 -m "Initial commit"
you can choose custom repo-name
(default: {algo}-{env_id}
) by passing a --repo-name
argument.
Download model from hub:
python -m utils.load_from_hub --algo ppo --env CartPole-v1 -f logs/ -orga sb3
The syntax used in hyperparameters/algo_name.yml
for setting hyperparameters (likewise the syntax to overwrite hyperparameters on the cli) may be specialized if the argument is a function. See examples in the hyperparameters/
directory. For example:
- Specify a linear schedule for the learning rate:
learning_rate: lin_0.012486195510232303
Specify a different activation function for the network:
policy_kwargs: "dict(activation_fn=nn.ReLU)"
We use Optuna for optimizing the hyperparameters. Not all hyperparameters are tuned, and tuning enforces certain default hyperparameter settings that may be different from the official defaults. See utils/hyperparams_opt.py for the current settings for each agent.
Hyperparameters not specified in utils/hyperparams_opt.py are taken from the associated YAML file and fallback to the default values of SB3 if not present.
Note: when using SuccessiveHalvingPruner ("halving"), you must specify --n-jobs > 1
Budget of 1000 trials with a maximum of 50000 steps:
python train.py --algo ppo --env MountainCar-v0 -n 50000 -optimize --n-trials 1000 --n-jobs 2 \
--sampler tpe --pruner median
Distributed optimization using a shared database is also possible (see the corresponding Optuna documentation):
python train.py --algo ppo --env MountainCar-v0 -optimize --study-name test --storage sqlite:///example.db
Print and save best hyperparameters of an Optuna study:
python scripts/parse_study.py -i path/to/study.pkl --print-n-best-trials 10 --save-n-best-hyperparameters 10
The default budget for hyperparameter tuning is 500 trials and there is one intermediate evaluation for pruning/early stopping per 100k time steps.
Note that the default hyperparameters used in the zoo when tuning are not always the same as the defaults provided in stable-baselines3. Consult the latest source code to be sure of these settings. For example:
-
PPO tuning assumes a network architecture with
ortho_init = False
when tuning, though it isTrue
by default. You can change that by updating utils/hyperparams_opt.py. -
Non-episodic rollout in TD3 and DDPG assumes
gradient_steps = train_freq
and so tunes onlytrain_freq
to reduce the search space.
When working with continuous actions, we recommend to enable gSDE by uncommenting lines in utils/hyperparams_opt.py.
We support tracking experiment data such as learning curves and hyperparameters via Weights and Biases.
The following command
python train.py --algo ppo --env CartPole-v1 --track --wandb-project-name sb3
yields a tracked experiment at this URL.
In the hyperparameter file, normalize: True
means that the training environment will be wrapped in a VecNormalize wrapper.
Normalization uses the default parameters of VecNormalize
, with the exception of gamma
which is set to match that of the agent. This can be overridden using the appropriate hyperparameters/algo_name.yml
, e.g.
normalize: "{'norm_obs': True, 'norm_reward': False}"
You can specify in the hyperparameter config one or more wrapper to use around the environment:
for one wrapper:
env_wrapper: gym_minigrid.wrappers.FlatObsWrapper
for multiple, specify a list:
env_wrapper:
- utils.wrappers.DoneOnSuccessWrapper:
reward_offset: 1.0
- sb3_contrib.common.wrappers.TimeFeatureWrapper
Note that you can easily specify parameters too.
You can specify which VecEnvWrapper
to use in the config, the same way as for env wrappers (see above), using the vec_env_wrapper
key:
For instance:
vec_env_wrapper: stable_baselines3.common.vec_env.VecMonitor
Note: VecNormalize
is supported separately using normalize
keyword, and VecFrameStack
has a dedicated keyword frame_stack
.
Following the same syntax as env wrappers, you can also add custom callbacks to use during training.
callback:
- utils.callbacks.ParallelTrainCallback:
gradient_steps: 256
You can specify keyword arguments to pass to the env constructor in the command line, using --env-kwargs
:
python enjoy.py --algo ppo --env MountainCar-v0 --env-kwargs goal_velocity:10
You can easily overwrite hyperparameters in the command line, using --hyperparams
:
python train.py --algo a2c --env MountainCarContinuous-v0 --hyperparams learning_rate:0.001 policy_kwargs:"dict(net_arch=[64, 64])"
Note: if you want to pass a string, you need to escape it like that: my_string:"'value'"
Record 1000 steps with the latest saved model:
python -m utils.record_video --algo ppo --env BipedalWalkerHardcore-v3 -n 1000
Use the best saved model instead:
python -m utils.record_video --algo ppo --env BipedalWalkerHardcore-v3 -n 1000 --load-best
Record a video of a checkpoint saved during training (here the checkpoint name is rl_model_10000_steps.zip
):
python -m utils.record_video --algo ppo --env BipedalWalkerHardcore-v3 -n 1000 --load-checkpoint 10000
Apart from recording videos of specific saved models, it is also possible to record a video of a training experiment where checkpoints have been saved.
Record 1000 steps for each checkpoint, latest and best saved models:
python -m utils.record_training --algo ppo --env CartPole-v1 -n 1000 -f logs --deterministic
The previous command will create a mp4
file. To convert this file to gif
format as well:
python -m utils.record_training --algo ppo --env CartPole-v1 -n 1000 -f logs --deterministic --gif
Final performance of the trained agents can be found in benchmark.md
. To compute them, simply run python -m utils.benchmark
.
List and videos of trained agents can be found on our Huggingface page: https://huggingface.co/sb3
NOTE: this is not a quantitative benchmark as it corresponds to only one run (cf issue #38). This benchmark is meant to check algorithm (maximal) performance, find potential bugs and also allow users to have access to pretrained agents.
7 atari games from OpenAI benchmark (NoFrameskip-v4 versions).
RL Algo | BeamRider | Breakout | Enduro | Pong | Qbert | Seaquest | SpaceInvaders |
---|---|---|---|---|---|---|---|
A2C | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
PPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
DQN | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
QR-DQN | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
Additional Atari Games (to be completed):
RL Algo | MsPacman | Asteroids | RoadRunner |
---|---|---|---|
A2C | ✔️ | ✔️ | |
PPO | ✔️ | ✔️ | |
DQN | ✔️ | ✔️ | |
QR-DQN | ✔️ | ✔️ |
RL Algo | CartPole-v1 | MountainCar-v0 | Acrobot-v1 | Pendulum-v1 | MountainCarContinuous-v0 |
---|---|---|---|---|---|
ARS | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
A2C | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
PPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
DQN | ✔️ | ✔️ | ✔️ | N/A | N/A |
QR-DQN | ✔️ | ✔️ | ✔️ | N/A | N/A |
DDPG | N/A | N/A | N/A | ✔️ | ✔️ |
SAC | N/A | N/A | N/A | ✔️ | ✔️ |
TD3 | N/A | N/A | N/A | ✔️ | ✔️ |
TQC | N/A | N/A | N/A | ✔️ | ✔️ |
TRPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
RL Algo | BipedalWalker-v3 | LunarLander-v2 | LunarLanderContinuous-v2 | BipedalWalkerHardcore-v3 | CarRacing-v0 |
---|---|---|---|---|---|
ARS | ✔️ | ✔️ | |||
A2C | ✔️ | ✔️ | ✔️ | ✔️ | |
PPO | ✔️ | ✔️ | ✔️ | ✔️ | |
DQN | N/A | ✔️ | N/A | N/A | N/A |
QR-DQN | N/A | ✔️ | N/A | N/A | N/A |
DDPG | ✔️ | N/A | ✔️ | ||
SAC | ✔️ | N/A | ✔️ | ✔️ | |
TD3 | ✔️ | N/A | ✔️ | ✔️ | |
TQC | ✔️ | N/A | ✔️ | ✔️ | |
TRPO | ✔️ | ✔️ |
See https://github.com/bulletphysics/bullet3/tree/master/examples/pybullet/gym/pybullet_envs.
Similar to MuJoCo Envs but with a free (MuJoCo 2.1.0+ is now free!) easy to install simulator: pybullet. We are using BulletEnv-v0
version.
Note: those environments are derived from Roboschool and are harder than the Mujoco version (see Pybullet issue)
RL Algo | Walker2D | HalfCheetah | Ant | Reacher | Hopper | Humanoid |
---|---|---|---|---|---|---|
ARS | ||||||
A2C | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
PPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
DDPG | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
SAC | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
TD3 | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
TQC | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
TRPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
PyBullet Envs (Continued)
RL Algo | Minitaur | MinitaurDuck | InvertedDoublePendulum | InvertedPendulumSwingup |
---|---|---|---|---|
A2C | ||||
PPO | ||||
DDPG | ||||
SAC | ||||
TD3 | ||||
TQC |
RL Algo | Walker2d | HalfCheetah | Ant | Swimmer | Hopper | Humanoid |
---|---|---|---|---|---|---|
ARS | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
A2C | ✔️ | ✔️ | ✔️ | |||
PPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | |
DDPG | ||||||
SAC | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
TD3 | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
TQC | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
TRPO | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
See https://gym.openai.com/envs/#robotics and DLR-RM#71
MuJoCo version: 1.50.1.0 Gym version: 0.18.0
We used the v1 environments.
RL Algo | FetchReach | FetchPickAndPlace | FetchPush | FetchSlide |
---|---|---|---|---|
HER+TQC | ✔️ | ✔️ | ✔️ | ✔️ |
See https://github.com/qgallouedec/panda-gym/.
Similar to MuJoCo Robotics Envs but with a free easy to install simulator: pybullet.
We used the v1 environments.
RL Algo | PandaReach | PandaPickAndPlace | PandaPush | PandaSlide | PandaStack |
---|---|---|---|---|---|
HER+TQC | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
To visualize the result, you can pass --env-kwargs render:True
to the enjoy script.
See https://github.com/maximecb/gym-minigrid A simple, lightweight and fast Gym environments implementation of the famous gridworld.
RL Algo | Empty | FourRooms | DoorKey | MultiRoom | Fetch |
---|---|---|---|---|---|
A2C | |||||
PPO | |||||
DDPG | |||||
SAC | |||||
TRPO |
There are 19 environment groups (variations for each) in total.
Note that you need to specify --gym-packages gym_minigrid
with enjoy.py
and train.py
as it is not a standard Gym environment, as well as installing the custom Gym package module or putting it in python path.
pip install gym-minigrid
python train.py --algo ppo --env MiniGrid-DoorKey-5x5-v0 --gym-packages gym_minigrid
This does the same thing as:
import gym_minigrid
You can train agents online using colab notebook.
We recommend using stable-baselines3 and sb3_contrib master versions.
apt-get install swig cmake ffmpeg
pip install -r requirements.txt
Please see Stable Baselines3 documentation for alternatives.
Build docker image (CPU):
make docker-cpu
GPU:
USE_GPU=True make docker-gpu
Pull built docker image (CPU):
docker pull stablebaselines/rl-baselines3-zoo-cpu
GPU image:
docker pull stablebaselines/rl-baselines3-zoo
Run script in the docker image:
./scripts/run_docker_cpu.sh python train.py --algo ppo --env CartPole-v1
To run tests, first install pytest, then:
make pytest
Same for type checking with pytype:
make type
To cite this repository in publications:
@misc{rl-zoo3,
author = {Raffin, Antonin},
title = {RL Baselines3 Zoo},
year = {2020},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/DLR-RM/rl-baselines3-zoo}},
}
If you trained an agent that is not present in the RL Zoo, please submit a Pull Request (containing the hyperparameters and the score too).
We would like to thank our contributors: @iandanforth, @tatsubori @Shade5 @mcres, @ernestum