[BUG] cannot use GymEnv() says '_set_gym_args' has not been found

[EkaterinaAbramova]

Describe the bug

The following is not working:

base_env = GymEnv("InvertedDoublePendulum-v4", device=device, frame_skip=frame_skip)
ModuleNotFoundError: Supported version of '_set_gym_args' has not been found.

To Reproduce

conda create --name gpu-torch-rl python=3.10 spyder=5.4.3
conda activate gpu-torch-rl

git clone https://github.com/pytorch/rl /Users/bondgirl007/anaconda3/repos/rl

cd /Users/bondgirl007/anaconda3/repos/rl/
pip install -e .

pip uninstall tensordict
pip install git+https://github.com/pytorch-labs/tensordict

Then I am trying to run this tutorial from PyTorch: https://pytorch.org/tutorials/intermediate/reinforcement_ppo.html
Everything up until this line is ok, but then the line doesn't run:
base_env = GymEnv("InvertedDoublePendulum-v4", device=device, frame_skip=frame_skip)

Expected behavior

to obtain base_env after calling GymEnv()

Screenshots

System info

MacBook Pro M2 max
Version: 13.5.2

import torchrl, numpy, sys
print(torchrl.__version__, numpy.__version__, sys.version, sys.platform)
0.1.1+18b33fe 1.26.0 3.10.13 (main, Sep 11 2023, 08:16:02) [Clang 14.0.6 ] darwin

[EkaterinaAbramova]

These seemed to have resolved the issue:
Supported version of '_set_gym_args' has not been found.
pip install gym --upgrade

DependencyNotInstalled: No module named 'mujoco'.
pip install mujoco

Sounds like something needs to be updated in the git download of torchrl, so that all of above doesn't happen for others.

[vmoens]

I believe this is due to a gym/gymnasium mismatch.
You can usually solve these issues with

with set_gym_backend("gym"):
    env = GymEnv(...)

that redirects all the decorated code to use gym (or gymnasium).
We had a thorough refactoring of our gym interactions in #1539 so now things should work more smoothly.

Feel free to reopen if the problem reappears!

[skandermoalla]

@vmoens I think @EkaterinaAbramova here just installed torchrl without any of the RL packages (gym, mujoco, etc.), and although this is a feature of TorchRL to not come with any of these dependencies, the first error message that a user encounters when creating a gym env is
ModuleNotFoundError: Supported version of 'torchrl.envs.libs.gym.GymEnv._set_gym_args' has not been found.
Which somehow hides the fact that the user is missing gym in their environment. Maybe this is an opportunity to catch that in _set_gym_args and improve its error message?

[vmoens]

Oh good catch thanks!
Indeed, we assume that people have installed everything in some weird places, it's true that the entry point of the gym classes isn't as obvious as I thought.