A training framework for Stable Baselines3 reinforcement learning agents, with hyperparameter optimization and pre-trained agents included.
WARNING: This version will be the last one supporting Python 3.6 (end of life in Dec 2021). We highly recommended you to upgrade to Python >= 3.7.
--load-last-checkpoint
(@SammyRamone)TypeError
for gym.Env
class entry points in ExperimentManager
(@schuderer)HER
is now a replay buffer class and no more an algorithmPlotNoiseRatioCallback
PlotActionWrapper
'lr'
key in Optuna param dict to 'learning_rate'
so the dict can be directly passed to SB3 methods (@justinkterry)utils.callbacks.ParallelTrainCallback
)--load-last-checkpoint
option for the enjoy scriptplotly
package required)scripts/plot_train.py
get_latest_run_id()
so it works in Windows too (@NicolasHaeffner)HER
replay bufferis_bullet()
to ExperimentManager
close()
for the enjoy scriptrequirements.txt
(@amy12xx)SAC
and TD3
search spacespanda-gym
environments (@qgallouedec)Blog post: https://araffin.github.io/post/sb3/
HER
handling action noiseHER
and enjoy scriptHER
hyperparametersLinearNormalActionNoise
sb3_contrib
is now requiredTimeFeatureWrapper
was moved to the contrib repoplot_train.py
script with updated plot_training_success.py
n_episodes_rollout
to train_freq
tuple to match latest version of SB3VecEnv
class to use for multiprocessingTQC
QR-DQN
from SB3 contribExperimentManager
classmake_env
with SB3 built-in make_vec_env
utils/utils.py
done)PPO
atari hyperparameters (removed vf clipping)A2C
atari hyperparameters (eps value of the optimizer)DQN
hyperparameters for CartPole