Closed shuds13 closed 5 months ago
Hi @shuds13, nice progress, this looks really promising. I have a couple of questions:
Optgen
is really the object in charge of carrying out the whole optimization (asking the generator for trials and launching them for evaluation). Within optimas
, this is basically what the Exploration
is supposed to do. So, I was wondering, could the run
method of Optgen
be moved to the Exploration
(with another name such as _libe_run
, as run
is already taken), and then we set gen_specs['gen_f']=exploration._libe_run
? This could also be done in a separate PR. Just wondering if passing a class method as gen_f
is possible (I think @jlnav mentioned this last meeting).
- Regarding the resources, the main thing we need is to be able to dedicate a GPU to the generator (or the manager, now). Is that something that can be done with this new approach?
My intention is that it will work just as it does now (as you say the manager will be assigned resources). I just haven't yet got around to ensuring the resources part works.
This is proof-of-principle for running the generator in-place on the manager.
This represents one possible approach (pros/cons)