[Feature request] Parsing train's arguments from a list of strings

Now that rl-baselines3-zoo is a package and the train function can be imported, I’d like to introduce the following feature

Current workflow

The learning is performed by calling a standalone python process such as:

python train.py <arguments>

Desired workflow

In an interactive programming environment such as jupyter notebooks I’d like it to behave as:

from rl_zoo3.train import train

argument_list: List[str] = [...]

train(argument_list)

This is possible because ArgumentParser’s method parse_args accepts a list of strings as input for parsing.

Changes to be introduced

A simple change on the behaviour of the train function should be enough, that is:

Option A:

rl_zoo3/train.py

import sys

def train(arguments_list: List[str] = sys.argv[1:]):
    ...
    args = parser.parse_args(arguments_list)

Option B:

rl_zoo3/train.py:

def train(arguments_list: List[str]):
    ...
    args = parser.parse_args(arguments_list)

train.py:

import sys
from rl_zoo3.train import train

if __name__ == "__main__":  # noqa: C901
    train(sys.argv[1:])

I’m more inclined to option B because it is much cleaner and straightforward to interpret what the program is doing.

Change purpose

When using jupyter notebook and the program raises an exception, the magic line %debug is used to inspect the stack frame, executing python train.py <args> forces us to use another method to live debug.

Final comments

All comments on why/why not the change should be done are always welcome. Also comments on how to improve my current workflow are welcome. Let me know if I can proceed with the PR.

[araffin]

Hello, you have actually a simple (slightly hacky solution), which allows you to do what you want:

import sys
from rl_zoo3.train import train

sys.argv = ["python", "--algo", "ppo", "--env", "Pendulum-v1"]

train()

[araffin]

so, different things: the zoo is not meant to be used from an interactive session (or only when debugging), if you need a more pythonic interface, you have the experiment manager and other utils available, or even SB3 api. Making the rl zoo a package was mainly to give access to the wrappers/callbacks/utils, and to be able to call the train/enjoy scripts from anywhere. I agree it’s not pretty but that’s only for debugging.