[Bug] pip install fails on atari-py

[aaronbannin]

🐛 Bug

Fresh install of extra from pip fails on atari-py.

UPDATE: Using the current master from git seems to work:
pip install git+https://github.com/DLR-RM/stable-baselines3#egg=stable-baselines3[extra]

To Reproduce

On a fresh install: pip install 'stable-baselines3[extra]' produces the following error:

WARNING: The candidate selected for download or install is a yanked version: 'atari-py' candidate (version 0.2.6 at https://files.pythonhosted.org/packages/43/dd/2721f34a89dc520d2e09363fd23d110a33bbab2399e50fdced6eb2ed2157/atari-py-0.2.6.tar.gz#sha256=6249ad5079b0489e87eb44e65485bb1b07cc1b5af729f1ee52ece749503ceb1d (from https://pypi.org/simple/atari-py/))
Reason for being yanked: re-release with new wheels
Building wheels for collected packages: atari-py
  Building wheel for atari-py (setup.py): started
  Building wheel for atari-py (setup.py): finished with status 'error'
  ERROR: Command errored out with exit status 1:
   command: ~/venv/bin/python3 -u -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'/private/var/folders/mw/fr9y9vy938b9pq99vq56rp000000gn/T/pip-install-4krkiw32/atari-py_89e87b2aae6b424c8b41a316c5a1417f/setup.py'"'"'; __file__='"'"'/private/var/folders/mw/fr9y9vy938b9pq99vq56rp000000gn/T/pip-install-4krkiw32/atari-py_89e87b2aae6b424c8b41a316c5a1417f/setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' bdist_wheel -d /private/var/folders/mw/fr9y9vy938b9pq99vq56rp000000gn/T/pip-wheel-40utx24i
       cwd: /private/var/folders/mw/fr9y9vy938b9pq99vq56rp000000gn/T/pip-install-4krkiw32/atari-py_89e87b2aae6b424c8b41a316c5a1417f/

If I try to install gym manually, based on this line, I get the following error:

Installing collected packages: tqdm, click, gym-notices, AutoROM.accept-rom-license, autorom, gym
  Attempting uninstall: gym
    Found existing installation: gym 0.19.0
    Uninstalling gym-0.19.0:
      Successfully uninstalled gym-0.19.0
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
stable-baselines3 1.4.0 requires gym<0.20,>=0.17, but you have gym 0.22.0 which is incompatible.

Expected behavior

Would love to see Successfully installed

### System Info

>>> sb3.get_system_info()
OS: macOS-12.1-x86_64-i386-64bit Darwin Kernel Version 21.2.0: Sun Nov 28 20:28:54 PST 2021; root:xnu-8019.61.5~1/RELEASE_X86_64
Python: 3.8.12
Stable-Baselines3: 1.4.0
PyTorch: 1.10.2
GPU Enabled: False
Numpy: 1.22.2
Gym: 0.22.0

({'OS': 'macOS-12.1-x86_64-i386-64bit Darwin Kernel Version 21.2.0: Sun Nov 28 20:28:54 PST 2021; root:xnu-8019.61.5~1/RELEASE_X86_64', 'Python': '3.8.12', 'Stable-Baselines3': '1.4.0', 'PyTorch': '1.10.2', 'GPU Enabled': 'False', 'Numpy': '1.22.2', 'Gym': '0.22.0'}, 'OS: macOS-12.1-x86_64-i386-64bit Darwin Kernel Version 21.2.0: Sun Nov 28 20:28:54 PST 2021; root:xnu-8019.61.5~1/RELEASE_X86_64\nPython: 3.8.12\nStable-Baselines3: 1.4.0\nPyTorch: 1.10.2\nGPU Enabled: False\nNumpy: 1.22.2\nGym: 0.22.0\n')

Additional context

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[aaronbannin]

Using the current master from git seems to work (with some fancy egg info):
git+https://github.com/DLR-RM/stable-baselines3#egg=stable-baselines3[extra]

[Miffyli]

Hmm this might simply be because stable-baselines3 pypi package has not been updated yet (as in: new release has not been made since gym update has been done). @araffin sounds like pretty serious issue, should probably do it asap when you have time?

[araffin]

Hello,

On a fresh install: pip install 'stable-baselines3[extra]' produces the following error:

I couldn't reproduce your error following the getting started colab... do you have the build essential package installed? (required for any compilation)

https://colab.research.google.com/github/Stable-Baselines-Team/rl-colab-notebooks/blob/sb3/stable_baselines_getting_started.ipynb

stable-baselines3 1.4.0 requires gym<0.20,>=0.17, but you have gym 0.22.0 which is incompatible.

SB3 1.4.0 requires gym < 0.20, so you should install gym 0.19 if you want to install it manually. And in fact, you need to install atari py manually because the correct version was not specified if I recall.

gym > 0.20 doesn't use atari py anymore...

[aaronbannin]

@araffin this was using Python installed on OS X, not Colab. I just reproduced with the instructions above on my work machine. This seems like an issue with PyPi propagation; an older version exists in PyPi but installing directly from Github gets the newest version?

python3 -m venv venv
source venv/bin/activate
pip install 'stable-baselines3[extra]'

[araffin]

an older version exists in PyPi but installing directly from Github gets the newest version?

which version are you talking about? SB3 version?
on PyPi, there is the latest stable release (1.4.0, compatible with gym 0.19) and since yesterday there is also the 1.4.1a1 which is compatible with gym 0.21 (and which uses ale-py instead of atari py)

See https://pypi.org/project/stable-baselines3/#history

[aaronbannin]

I don't know why pip install 'stable-baselines3[extra]' fails and pip install git+https://github.com/DLR-RM/stable-baselines3#egg=stable-baselines3[extra] succeeds.

[araffin]

could you try pip install 'stable-baselines3[extra]==1.4.1a1' ?

[aaronbannin]

Yup!

Successfully installed AutoROM.accept-rom-license-0.4.2 absl-py-1.0.0 ale-py-0.7.4 autorom-0.4.2 cachetools-5.0.0 certifi-2021.10.8 charset-normalizer-2.0.12 click-8.0.4 cloudpickle-2.0.0 cycler-0.11.0 fonttools-4.29.1 google-auth-2.6.0 google-auth-oauthlib-0.4.6 grpcio-1.44.0 gym-0.21.0 idna-3.3 importlib-metadata-4.11.1 importlib-resources-5.4.0 kiwisolver-1.3.2 markdown-3.3.6 matplotlib-3.5.1 numpy-1.22.2 oauthlib-3.2.0 opencv-python-4.5.5.62 packaging-21.3 pandas-1.4.1 pillow-9.0.1 protobuf-3.19.4 psutil-5.9.0 pyasn1-0.4.8 pyasn1-modules-0.2.8 pyparsing-3.0.7 python-dateutil-2.8.2 pytz-2021.3 requests-2.27.1 requests-oauthlib-1.3.1 rsa-4.8 six-1.16.0 stable-baselines3-1.4.1a1 tensorboard-2.8.0 tensorboard-data-server-0.6.1 tensorboard-plugin-wit-1.8.1 torch-1.10.2 tqdm-4.62.3 typing-extensions-4.1.1 urllib3-1.26.8 werkzeug-2.0.3 wheel-0.37.1 zipp-3.7.0

[araffin]

As we cannot update SB3 1.4.0 pypi package and that the new version works, I guess the issue is solved?

[aaronbannin]

If anyone tries a simple pip install, it will still fail, right? Perhaps a version bump would provide stability? If 1.4.1a1 got promoted to 1.5.0?

[araffin]

If anyone tries a simple pip install, it will still fail, right?

Only on Mac OS with your setup when using the [extra] argument (where something goes wrong at compile time apparently).

erhaps a version bump would provide stability? If 1.4.1a1 got promoted to 1.5.0?

I would like to wait a bit before doing a new release, mostly to catch potential bugs and check full compatibility with the newest gym version (#780 requires additional work).
In the meantime, one can always install from source or use the alpha version on pypi.

[olenmg]

If anyone tries a simple pip install, it will still fail, right? Perhaps a version bump would provide stability? If 1.4.1a1 got promoted to 1.5.0?

Hello. I'm not sure if the stable-baselines3 fully supports M1 mac, but I also failed with pip install and succeed with 1.4.1a1 on my m1 pro mac.

[araffin]

stable-baselines3 fully supports M1 mac

stable-baselines3 is a python only package, so as long as you manage to install pytorch, you can use SB3. The issue is with atari-py from the [extra] requirements. As M1 is an arm processor (vs intel processor usually), some pre-built binaries may be missing.
@aaronbannin I think this is your issue too. And because ale-py comes with pre-built binaries for arm, the new SB3 version (1.4.1a1) works with M1 too ;)

[aaronbannin]

@araffin I'm on an Intel chip. I think the core problem might be the PyPi artifact has a dependency graph that points to a version of atari-py that was removed from PyPi.