Including stable-baselines3 and torch & cuda dependencies in requirements

[hagrid67]

We have introduced stable-baselines3 into the dependencies - see pyproject.toml and the generated requirements.txt and requirements-dev.txt. This package requires torch which in turn requires various nvidia packages; this makes the virtual env several Gb in size. Previously we have separated ML-related libraries so that Flatland itself is a relatively lightweight and ML-agnostic package.

Are we happy with this change or do we need to discuss further?

(edit title from reqs-dev to reqs because currently baselines3 appears in reqs)

[fnberta]

As @manuschn mentioned on slack, no new dependencies were introduced when we moved to pyproject.toml. What we did however is make all the existing ones explicit. Previously, there were a lot of dependencies not listed anywhere. I did not special case the contrib folder in any way though. Maybe it is something special and you want to handle dependencies differently? I'm also seeing just now that there is a requirements_training.txt in there. Not sure what role this should play. In that sense, thanks for opening the discussion. @manuschn @aiAdrian what are your thoughts on this?

[hagrid67]

Yeah not sure the best way to handle it in the CI structure but I'd be inclined to make another package(s) which includes flatland and additionally the heavyweight GPU/ML packages (which FL itself doesn't yet use). So training, baselines, solutions etc.

I think we had something like that in the past but I'm not familiar with the structure we used, other than contrib.

[hagrid67]

Currently the stables-baselines3 appears in the main requirements; IMO it should at least be demoted to reqs-dev, and should be in reqs-training if that is available.

I take your point that it was always a latent dependency; but running the pettingzoo baselines was a fairly niche activity; someone doing that could easily pip install torch etc; it did not need to be an explicit dependency which affected everyone trying to just install flatland.

flatland itself is like one of the many small environments in gym / pettingzoo: it doesn't know anything about training etc; it's just a small program with the capability to generate a graphical / visual output. The packages for this are relatively small (a few kb to a few Mb, numpy is 36Mb)

The flatland dev requirements are for developing flatland; this could be seen as just developing the environment. This includes some fairly bulky things like jupyter (~20Mb across a few packages?)

Do we have the facility for separate requirements-training, as well as requirements-dev? (I got the impression there was just a single bit switch to enable "optional-dependencies" in pyproject.toml, so the answer would appear to be no. But if we do, then we could push all the GPU / ML packages and the packages which need them into reqs-training

The GPU / ML libraries (eg torch ~1.3Gb, nvidia ~2.6Gb) are an order of magnitude or two bigger than packages like numpy (36Mb).

FYI Some of the larger package sizes from the virtual-env I'm using:
13M networkx
13M rapidfuzz
15M cryptography
15M pip
16M jupyterlab
17M wandb
18M debugpy
18M lxml
24M botocore
34M babel
35M scipy.libs
36M numpy
37M numpy.libs
38M matplotlib
42M nbclassic
52M sympy
61M pandas
72M cmake
76M panel
85M bokeh
86M scipy
115M pyarrow
162M ray
185M triton
1.3G torch
2.6G nvidia

(created using the command du -sh * | sort -h in the site-packages directory; this is from my general virtual env which may contain stuff not required by flatland)

[hagrid67]

Meh - I've moved stable_baselines3 from the dependencies into the optional dependencies so that torch and nvidia are no longer core requirements but just appear in reqs-dev.

[aiAdrian]

Great hagrid67 ***@***.***> schrieb am Mi., 6. Dez. 2023, 10:00:

…

Meh - I've moved stable_baselines3 from the dependencies into the optional dependencies so that torch and nvidia are no longer core requirements but just appear in reqs-dev. — Reply to this email directly, view it on GitHub <#63 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACPQ2XTXPCHYHTKNBN5BJNTYIAX4LAVCNFSM6AAAAABAE2SSQOVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM3TONZTHA3TO> . You are receiving this because you were mentioned.Message ID: <flatland-association/flatland-rl/repo-discussions/63/comments/7773877@ github.com>