Split mlagents into two packages

[awjuliani]

This PR splits the mlagents package into to packages. One is mlagents_envs which contains only the python api, and is nice and lightweight. The second is mlagents which contains the PPO/BC trainer and depends on mlagents_envs.

[vincentpierre]

any reason for this change ?

[awjuliani]

This is the new latest, so it is more likely that people on mac won't have to change the name here.

[vincentpierre]

Is it compatible with grpcio>=1.11.0,<1.12.0 on all 3 platforms ?

[awjuliani]

It worked for me.

[ervteng]

Linux - works as well.

[ervteng]

I think we also should change the documentation (Installation.md) to have users install both ml-agents and ml-agents-envs. Once we push mlagents_envs to pypi, pip install -e . on ml-agents will work, but it might be confusing to people, as it'll pull a version that is independent of the repo they just cloned.

Otherwise functionally it seems to work OK!

[awjuliani]

@ervteng Maybe we should tell people to just use the pypi packages, instead of installing from the repo source? That may be more straightforward for most users.

I also realized that we should test mlagents_envs on a python 3.7 environment, to ensure that it works there. If it does, we should change the requirements to allow it.

[harperj]

@awjuliani looks like this needs to be rebased for us to include it in v0.8 and tests fixed.

[ervteng]

Also I realized we need to change the docs (the Readme in the ml-agents directory) to reflect the change.

[ervteng]

@ervteng Maybe we should tell people to just use the pypi packages, instead of installing from the repo source? That may be more straightforward for most users.

We might be able to have our cake and eat it too... So for RLLib they use something called "extras" in the setup file. Example:

setup(
    name="Project-A",
    ...
    extras_require={
        'PDF':  ["ReportLab>=1.2", "RXP"],
        'reST': ["docutils>=0.3"],
    }
)

PDF and ReST are "extras" in this package (ours would be trainers, with envs being non-extra). We can specify different requirements for each. In this way, installing ml-agents would install just the envs package, and doing pip install ml-agents[trainers] would install both.

You could install using pip install -e ./ [trainers] as well. This way we'll have just one setup file and we won't need to change the directory structure or even the code's import statements.

[vincentpierre]

This file should not be here, it should be with the others in ml-agents/ml-agents-envs/mlagents/envs/communicator_objects

[ervteng]

You mean the ones in ml-agents/mlagents right? I removed those three - but I think this one should be here.