Timeline for Batch wise Inference to speed up Muzero's MCTS
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 5 at 23:52 | comment | added | KarelPeeters | I'm happy to hear that! Yeah using a systems language for tree search and python for training is pretty common, LC0 and KataGo are doing the same thing. Indeed for AlphaZero selfplay is the part that's important to optimize, training is only a small fraction of the total cost. | |
| Mar 5 at 21:45 | comment | added | Lynix | This is has helped me a lot. I really like the fact that you use something fast like rust for the mcts and just use python for the nn queries and training. For now im just using the library ray to run games in parallel. Its still pretty slow. For connect4 it takes 7min to run 100 games with 50 simulations per move. The training itself is pretty fast. I've implemented batchwise unrolling to generate predictions. I'm able to do 30 training steps per second at batch size 32. | |
| Mar 5 at 21:33 | vote | accept | Lynix | ||
| Feb 17 at 12:16 | history | edited | KarelPeeters | CC BY-SA 4.0 |
added 3 characters in body
|
| Feb 14 at 23:18 | history | answered | KarelPeeters | CC BY-SA 4.0 |