[question] [feature request] DDPG VecEnv support

[Enderdead]

I'm currently working to speed up a project using the stable-baselines DDPG algorithm. My environment can be easily converted into VecEnv. But I figured out that your DDPG doesn't support the VecEnv.
I'm wondering why you didn't implement it? Lack of time, hard to deal with?
Thank you in advance.

[araffin]

Hello,

Duplicate of #170 (related to #495 )

In short, SAC/TD3/DDPG were designed for using them on real robot, so where multiprocessing is not possible. I would recommend using PPO/A2C for fast training using massive multiprocessing (where sample efficiency is not a problem but wall clock time is)

In theory, SAC/TD3/DDPG could be extended to multiprocessing (this will maybe easier to do so with v3 #576 ), however, in practice, this would complicate the current code (maybe a wrapper around the replay buffer would be a solution).

Perhaps this is solved with #684 ?

[araffin]

Perhaps this is solved with #684 ?

Only partially solved, we would need to change all the logging around (this would complexify the code a bit) and for some algorithm (like DDPG), we are using a UnvecWrapper (for legacy reason, not planned in the next version) to have only one environment, so switching to a VecEnv may break other things unfortunately.