search

ucl-dark / pax

Scalable Opponent Shaping Experiments in JAX
Apache License 2.0
21 stars 5 forks source link

Proper installation guidelines #119

Open samvelyan opened 2 years ago

samvelyan commented 2 years ago

The users should be able to run the framework following ONLY the installation guide.

I think easiest to use via conda.

akbir commented 1 year ago

hard veto on conda - i think this is all pip installable!

akbir commented 1 year ago

like we don’t need auto install binaries so feels like conda is overkill

akbir commented 1 year ago

@newtonkwan - can you update the install details and look at making pax a package?

samvelyan commented 1 year ago

what does pip-installable have to do with having conda environments for your py projects?

akbir commented 1 year ago

So the way I see it, conda is a system package manager vs pip is a python package manager.

You should use conda (or suggest it to users) when you need to install things that are not just in python. For example - torch or opencv, where the dependencies are C++ binaries or other libraries (e.g. libjpeg-turbo).

Just like most other packages in the Jax eco-system you don't need to specify any additional binaries (e.g evosax, Chex, rlax) - the user is assumed to have jax already installed. I think it would be good to keep to that standard.