Stable Baselines3 (SB3) is a set of reliable implementations of reinforcement learning algorithms in PyTorch. It is the next major version of Stable Baselines.
Github repository: https://github.com/DLR-RM/stable-baselines3
Paper: https://jmlr.org/papers/volume22/20-1364/20-1364.pdf
RL Baselines3 Zoo (training framework for SB3): https://github.com/DLR-RM/rl-baselines3-zoo
RL Baselines3 Zoo provides a collection of pre-trained agents, scripts for training, evaluating agents, tuning hyperparameters, plotting results and recording videos.
SB3 Contrib (experimental RL code, latest algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
Main Features
- Unified structure for all algorithms
- PEP8 compliant (unified code style)
- Documented functions and classes
- Tests, high code coverage and type hints
- Clean code
- Tensorboard support
pip install godot-rl[sb3]gdrl.env_from_hub -r edbeeching/godot_rl_<ENV_NAME>
chmod +x examples/godot_rl_<ENV_NAME>/bin/<ENV_NAME>.x86_64 # linux exampleClone the repository or download the script sb3 example. To use the example script, first move to the location where the downloaded script is in the console/terminal, and then try some of the example use cases below:
python stable_baselines3_example.pypython stable_baselines3_example.py --env_path=path_to_executableFor the previously downloaded envs, the path will be e.g.
--env_path=examples/godot_rl_<ENV_NAME>/bin/<ENV_NAME>.x86_64
Note that the exported environment will not be rendered in order to accelerate training.
If you want to display it, add the --viz argument.
python stable_baselines3_example.py --env_path=path_to_executable --n_parallel=4python stable_baselines3_example.py --env_path=path_to_executable --speedup=8You can optionally set an experiment directory and name to override the default. When saving checkpoints, you need to use a unique directory or name for each run (more about that below).
python stable_baselines3_example.py --experiment_dir="experiments" --experiment_name="experiment1"The exported .onnx model can be used by the Godot sync node to run inference from Godot directly, while the saved .zip model can be used to resume training later or run inference from the example script by adding --inference.
python stable_baselines3_example.py --timesteps=100_000 --onnx_export_path=model.onnx --save_model_path=model.zipNote: If you interrupt/halt training using ctrl + c, it should save/export models before closing training (but only if you have included the corresponding arguments mentioned above). Using checkpoints (see below) is a safer way to keep progress.
This will load the previously saved model.zip, and resume training for another 100 000 steps, so the saved model will have been trained for 200 000 steps in total.
Note that the console log will display the total_timesteps for the last training session only, so it will show 100000 instead of 200000.
python stable_baselines3_example.py --timesteps=100_000 --save_model_path=model_200_000_total_steps.zip --resume_model_path=model.zipYou can save periodic checkpoints and later resume training from any checkpoint using the same CL argument as above, or run inference on any checkpoint just like with the saved model.
Note that you need to use a unique experiment_name or experiment_dir for each run so that checkpoints from one run won't overwrite checkpoints from another run.
Alternatively, you can remove the folder containing checkpoints from a previous run if you don't need them anymore.
E.g. train for a total of 2 000 000 steps with checkpoints saved at every 50 000 steps:
python stable_baselines3_example.py --experiment_name=experiment1 --timesteps=2_000_000 --save_checkpoint_frequency=50_000Checkpoints will be saved to logs\sb3\experiment1_checkpoints in the above case, the location is affected by --experiment_dir and --experiment_name.
You can run inference on a model that was previously saved using either --save_model_path or --save_checkpoint_frequency.
python stable_baselines3_example.py --timesteps=100_000 --resume_model_path=model.zip --inferenceBy default, the learning rate will be constant throughout training.
If you add --linear_lr_schedule, learning rate will decrease with the progress,
and reach 0 at --timesteps value.
python stable_baselines3_example.py --timesteps=1_000_000 --linear_lr_scheduleIf you want to report success rate based on some condition (e.g. whether the agent successfully finished the level or not), follow the steps below:
var is_success := false
func get_info() -> Dictionary:
if done:
return {"is_success": is_success}
return {}The above snippet will send the information on whether or not the episode was succesful to the Python training server. SB3 can use this to report the success rate.
The condition depends on your use case, for example, here's how we can implement this in the SimpleReachGoal env.
In the player.gd script, we just add is_success to depend on whether or not the reward is higher than 0:
## Ends the game, setting an optional reward
func game_over(reward: float = 0.0):
ai_controller.is_success = reward > 0
ai_controller.reward += reward
game_scene_manager.reset()Notes:
- Although not directly visible, the
donecondition is also set by this method (by callinggame_scene_manager.reset()), in a different env it might be something such as:
func game_over():
ai_controller.is_success = reward > 0
ai_controller.done = true
ai_controller.needs_reset = true- The condition for success can vary based on your environment, it does not have to depend directly on the reward.
- The current sb3 docs relevant section suggests:
you must pass an extra argument to the Monitor wrapper to log that value (info_keywords=("is_success",)
We didn't add this to the SB3 example script since it seems to work without the value in the current SB3 version, as we didn't test this in-depth yet - try adding the argument in case of any issues.
After these changes, you should be able to see the rate in the training stats, e.g.:
You can see the output from the training session in tensorboard. Check this guide for more info.