Merge pull request #136 from edbeeching/sb3_example_docs_update

Update ADV_STABLE_BASELINES_3.md to add instructions on using the command line arguments with the example file.

docs/ADV_STABLE_BASELINES_3.md

@@ -46,42 +46,66 @@ While the default options for sb3 work reasonably well. You may be interested in
 
 We recommend taking the [sb3 example](https://github.com/edbeeching/godot_rl_agents/blob/main/examples/stable_baselines3_example.py) and modifying to match your needs. 
 
-This example exposes more parameter for the user to configure, such as `--speedup` to run the environment faster than realtime and the `--n_parallel` to launch several instances of the game executable in order to accelerate training (not available for in-editor training).
+The example exposes more parameters for the user to configure, such as `--speedup` to run the environment faster than realtime and the `--n_parallel` to launch several instances of the game executable in order to accelerate training (not available for in-editor training). 
 
+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
-import argparse
-
-from godot_rl.wrappers.stable_baselines_wrapper import StableBaselinesGodotEnv
-from stable_baselines3 import PPO
+### Train a model in editor:
+```bash
+python stable_baselines3_example.py
+```
 
-# To download the env source and binary:
-# 1.  gdrl.env_from_hub -r edbeeching/godot_rl_BallChase
-# 2.  chmod +x examples/godot_rl_BallChase/bin/BallChase.x86_64
+### Train a model using an exported environment:
+```bash
+python stable_baselines3_example.py --env_path=path_to_executable
+```
+Note that the exported environment will not be rendered in order to accelerate training.
+If you want to display it, add the `--viz` argument.
 
+### Train an exported environment using 4 environment processes:
+```bash
+python stable_baselines3_example.py --env_path=path_to_executable --n_parallel=4
+```
 
-parser = argparse.ArgumentParser(allow_abbrev=False)
-parser.add_argument(
-    "--env_path",
-    # default="envs/example_envs/builds/JumperHard/jumper_hard.x86_64",
-    default=None,
-    type=str,
-    help="The Godot binary to use, do not include for in editor training",
-)
+### Train an exported environment using 8 times speedup:
+```bash
+python stable_baselines3_example.py --env_path=path_to_executable --speedup=8
+```
 
-parser.add_argument("--speedup", default=1, type=int, help="whether to speed up the physics in the env")
-parser.add_argument("--n_parallel", default=1, type=int, help="whether to speed up the physics in the env")
+### Set an experiment directory and name:
+You 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).
+```bash
+python stable_baselines3_example.py --experiment_dir="experiments" --experiment_name="experiment1"
+```
 
-args, extras = parser.parse_known_args()
+### Train a model for 10_000 steps then save and export the model
+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`.
+```bash
+python stable_baselines3_example.py --timesteps=100_000 --onnx_export_path=model.onnx --save_model_path=model.zip
+```
 
+### Resume training from a saved .zip model
+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`. 
+```bash
+python stable_baselines3_example.py --timesteps=100_000 --save_model_path=model_200_000_total_steps.zip --resume_model_path=model.zip
+```
 
-env = StableBaselinesGodotEnv(env_path=args.env_path, show_window=True, n_parallel=args.n_parallel, speedup=args.speedup)
+### Save periodic checkpoints
+You 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.
 
-model = PPO("MultiInputPolicy", env, ent_coef=0.0001, verbose=2, n_steps=32, tensorboard_log="logs/sb3")
-model.learn(200000)
+E.g. train for a total of 2 000 000 steps with checkpoints saved at every 50 000 steps:
 
-print("closing env")
-env.close()
+```bash
+python stable_baselines3_example.py --experiment_name=experiment1 --timesteps=2_000_000 --save_checkpoint_frequency=50_000
+```
 
+Checkpoints will be saved to `logs\sb3\experiment1_checkpoints` in the above case, the location is affected by `--experiment_dir` and `--experiment_name`.
 
-```
+### Run inference on a saved model for 100_000 steps
+You can run inference on a model that was previously saved using either `--save_model_path` or `--save_checkpoint_frequency`.
+```bash
+python stable_baselines3_example.py --timesteps=100_000 --resume_model_path=model.zip --inference
+```