Training infrastructure
Closed this issue · 4 comments
Hi,
I was wondering if there is any guideline regarding the suggested infrastructure to run the training using this library.
I've read this but would be great to have more info! I was thinking to use GKE but not really sure how to make the setup.
Thanks!
It is easiest to just train one policy on one Linux machine, with many cpu cores. Doesn't matter if that is a local desktop or google cloud instance or whatever. Install is trivial as in the readme
ok thanks. I'm running a simulation on my 8 cores ubuntu machine. After 1.5h, around ~1M step were computed, so it should takes around 6 days to reach 100M steps. I probably need a better machine :)
Indeed. I'm using the AMD Ryzen 3900, with 12 cores/24 threads, takes 3 days for 200Mln samples. I would go for AMD 39?0x threadripper nowadays, dependent on budget. Hopefully we find some faster learning method, perhaps based on the new tiny differentiable simulator (TDS).
Time to buy a new machine. thanks for your setup suggestions!