[2023 January] Changes required to run installation on macosx

[nirajpandkar]

1. Dependency upgrades in setup.py
2. tensorflow api change from tf.train to tf.compat.v1.train

If you are trying to follow these instructions to install - https://spinningup.openai.com/en/latest/user/installation.html, I had to do these changes in order to get it working on a MacBook pro 2019 version.

PS. Additional dependent libraries also required to successfully run the training (other than the instructions given on the installation page)

1. pip install swig
2. brew install freetype
3. Command Line Tools for xcode: xcode-select --install (The dialog box says it'll take 200+ hours to complete so install it externally - https://developer.apple.com/download/all/?q=command)

[reaktor-esakoskinen]

I arrived at almost exactly the same changes trying to install on an 2022 M1 Pro. However, since an M1-compatible TensorFlow wheel is not distributed through the normal Anaconda repositories, I had to take the additional steps documented here:
tensorflow/tensorflow#51506 (comment)

In order to get OpenGL rendering to work, I also needed pip install pyglet==1.5.15.

These might be useful to mention when the Gym's install instructions are updated.

There was also another PR with more comprehensive changes for TensorFlow 2 compatibility, but so far I didn't need them for running experiments:
#341

[KaiOnGitHub]

I additionally had to change tensorflow to tensorflow-macos<3.0 to support apple's GPU supported Tensorflow version.

Here are all commands that yielded a successful install for me (I already had the command line tools and used pipenv to create the python environment)

git clone https://github.com/openai/spinningup.git
cd spinningup
git fetch origin pull/381/head:2023Jan_dependency_upgrades
git checkout 2023Jan_dependency_upgrades
pipenv --python 3.9
pipenv shell
pip install swig
brew install freetype
pip install pyglet==1.5.15
pip install tensorflow-macos
pip install tensorflow-metal

replace 'tensorflow>=1.8.0,<3.0 in setup.py with:
'tensorflow>=1.8.0,<3.0; sys_platform != "darwin" or platform_machine != "arm64"',
'tensorflow-macos>=1.8.0,<3.0; sys_platform == "darwin" and platform_machine == "arm64"',

Finally run the install:
pip install -e .

I somehow had to run pip install pyglet==1.5.15 again.

[rojas70]

Do we have the equivalent for linux ubuntu?

[Nos-Talion]

I am sorry that I followed the instructions and successfully installed the spinningup but when I use the command python -m spinup.run test_policy data/installtest/installtest_s0, I got zsh: illegal hardware instruction python -m spinup.run test_policy data/installtest/installtest_s0. Could someone help me?

[mrvincenzo]

worked like a charm on M1 with python 3.8.

1. Apply changes from PR File Changes
2. conda create -n spinningup python=3.8
3. pip install -e .

More issues arise later with

python -m spinup.run plot
python -m spinup.run test_policy

after successful run of

python -m spinup.run ppo --hid "[32,32]" --env LunarLander-v2 --exp_name installtest --gamma 0.999

How to fix python -m spinup.run plot ...?

1. in setup.py replace
   'matplotlib==3.1.1' -> 'matplotlib==3.7.5'
'seaborn==0.8.1' -> 'seaborn==0.13.2'

2. Replace

sns.tsplot(data=data, time=xaxis, value=value, unit="Unit", condition=condition, ci='sd', **kwargs)

with

sns.lineplot(data=data, x=xaxis, y=value, hue=condition, ci="sd", **kwargs)

How to fix python -m spinup.run test_policy ...?

pip install pyglet==1.5.29

you will get an error message

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.
gym 0.15.7 requires pyglet<=1.5.0,>=1.4.0, but you have pyglet 1.5.29 which is incompatible.

but the installation should succeed (verify with pip show pyglet)

[Nos-Talion]

Sorry, I tried but it still showed zsh: illegal hardware instruction tonymontana ***@***.***> 于2024年5月25日周六 00:41写道：

…

worked like a charm on M1 with python 3.8. 1. Apply this PR File Changes <https://github.com/openai/spinningup/pull/381/files> 2. conda create -n spinningup python=3.8 3. pip install -e . — Reply to this email directly, view it on GitHub <#381 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQO2BAMTYHIHOJMMZG5R3PDZD5UTFAVCNFSM6AAAAAAUCVWOYOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMRZHE3TCMBRHA> . You are receiving this because you commented.Message ID: ***@***.***>

[takeshi8989]

worked like a charm on M1 with python 3.8.

1. Apply changes from PR File Changes
2. conda create -n spinningup python=3.8
3. pip install -e .

More issues arise later with

python -m spinup.run plot
python -m spinup.run test_policy

after successful run of

python -m spinup.run ppo --hid "[32,32]" --env LunarLander-v2 --exp_name installtest --gamma 0.999

How to fix python -m spinup.run plot ...?

1. in setup.py replace
   'matplotlib==3.1.1' -> 'matplotlib==3.7.5'
'seaborn==0.8.1' -> 'seaborn==0.13.2'
2. Replace

sns.tsplot(data=data, time=xaxis, value=value, unit="Unit", condition=condition, ci='sd', **kwargs)

with

sns.lineplot(data=data, x=xaxis, y=value, hue=condition, ci="sd", **kwargs)

How to fix python -m spinup.run test_policy ...?

pip install pyglet==1.5.29

you will get an error message

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.
gym 0.15.7 requires pyglet<=1.5.0,>=1.4.0, but you have pyglet 1.5.29 which is incompatible.

but the installation should succeed (verify with pip show pyglet)

I also had to run
pip install swig
And then
pip install -e .