Comparison with tensorflow1/2 implementation

[a-z-e-r-i-l-a]

I wanted to ask if you expect better performance using pytorch in stable-baselines3 than if it was going to be implemented with tensorflow2 ( or compared to the currently implemented tensorflow1). If so, from what aspects do you expect the advantages be in the end?

[Miffyli]

Possibly some related discussion in #20

It is hard to say yet how far the optimizations can go, but we expect slightly lower performance (in terms of wall-clock time) than TF1 because PyTorch is eager. However this should not be significant after code is written to follow PyTorch's guidelines, and it will provide much cleaner code overall. We chose PyTorch over TF2 partially because of majority vote.

[araffin]

Hello,
what type of performance are you talking about?
wall clock time? score in the environments?

[a-z-e-r-i-l-a]

yea I meant about wall clock time or any other aspect which you may have thought about, specifically compared to TF2.

It seemed to me that there is a tendency towards PyTorch more than TF2 but I couldn't guess why. and was wondering if there is any advantage in using PyTorch compared to TF2.

[araffin]

i guess reading those two issues will give you the answer then:

• V3 new backend: PyTorch? and the future of Stable Baselines hill-a/stable-baselines#733
• V3.0 implementation design hill-a/stable-baselines#576