-
Notifications
You must be signed in to change notification settings - Fork 724
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add minimal TF2 support #1026
base: master
Are you sure you want to change the base?
Add minimal TF2 support #1026
Conversation
I would rather keep a separate branch for minimal tf2 support, as this requires tf>=1.15. Regarding tf-contrib, all the issues are: https://github.com/Stable-Baselines-Team/stable-baselines/search?q=contrib |
I see your points and agree this could be a separate branch, but we could inform users somehow that they should install the branch version (e.g. warning if they are importing stable-baselines with TF2.x?). As it stands stable-baselines will die once support for TF1 ends in terms of outdated libraries and whatnot, and there is bunch of research code that might rely on stable-baselines in this current form rather than stable-baselines3 and such. |
yes, in the doc and readme. For the
well, we don't plan to support SB2 forever anyway, no? |
Ok. I was thinking of cases where somebody installs TF2 afterwards, but to be fair that one is going to be on them (should follow what dependencies say).
Oh no, can not assume so :) . Given SB2's popularity I just would like to see a tiny bit more extension to its life-line. |
any progress on this? |
this should answer your question: #366 SB3 repo (pytorch): https://github.com/DLR-RM/stable-baselines3 SBX (jax, experimental): https://github.com/araffin/sbx SB2 tf2 (unnofficial): https://github.com/sophiagu/stable-baselines-tf2 SB tf2 (experimental, archive): https://github.com/Stable-Baselines-Team/stable-baselines-tf2/ |
thank you! |
As per title, add minimal TF2 support while keeping TF1 support.
Motivation and Context
Closes #1012
Also motivated by the fact that TF2 offers relatively simple backwards-compatibility with TF1-like code, for now.
As noted in this comment, most of the change is just about replacing that one import. The second big thing is
tensorflow.contrib
, but updating for it seems reasonably doable. I am mostly worried about the hidden changes to agent performance and/or running time.TODO
tensorflow.compat.v1
to import TF1-like tensorflow.tensorflow.contrib
(mainlytensorflow.contrib.layers
).Types of changes
Checklist:
pytest
andpytype
both pass (by runningmake pytest
andmake type
).