Michael Pang
Michael Pang
Hi, nice project! It's been very useful to us working on the chess version. Please feel free to use any part of my code if it's helpful. I did implement...
I did get a significant speedup somehow, although it might be because I rewrote almost the entire player.py and made a bunch of subconscious optimizations along the way. It's hard...
Well for one I believe LZ uses a 5x64 model while we're using 7x256, and most of the FLOPs are in the conv layers, not the input layer. LZ may...
Yea you're absolutely right about the 19x19, I missed that part. RAZ has 7 residual blocks while LZ has 5, and FLOPs is roughly proportional to number of blocks. (coincidentally,...
Have you tried the tensorflow version? (lc0 folder)
If I understand correctly this issue is more about time management than puct tuning. It proposes dynamically increasing the number of playouts based on the stability of the search, which...
This is because it passes the color to move as an input plane, so you get slightly different outputs. It should really be removed (IMO), but it was decided a...
It seems that leela's main weakness atm is the endgame, but wouldn't a temperature scheme like this tend to reach the endgame less often?
Perhaps one thing to consider is that the history planes encode some information about the opponent, which isn't really part of "pure" chess but tends to be something humans use...
Sure, I'll give it a shot. I'm still new to this so could you tell me what the best practice is for taking multiple file names in a command line...