Open pangafu opened 6 years ago
Hi, just a note that I'm the one running LZ-1a198f-t1-v2600
. I guess you are running LZ-H6-W27-V2600
? I posted a little about my plans here: https://www.reddit.com/r/cbaduk/comments/823ekq/lz1a198ft1v2600_on_cgos/
@killerducky yes. I now start training a 20*128 network and hybrid for some test, now not run on CGOS.
https://github.com/gcp/leela-zero/issues/954
To see the true power of the hybrid, I hybrid 2 weights and run it on cgos.
W23 is inherited from No.86 weight (9ac646d7) and other 5 weight after it , W24 is inherited from No.86(9ac646d7) and No.87(92dd0397) and other 7 weights between them. And you can see the PO 200 match log in the zip file.
And there is many other leela zero weights run on cgos, such as LZ-0xx-p1600-t1-r1, StellarLZ-p6400, LeelaZero-test... I think LZ-HY-W23-PO1600 's parameter is same as LZ-0xx-p1600-t1-r1, so the result of them can compare.
LZ-HY-W24-PO1600 3030
LZ-HY-W23-PO1600 3011
And other LZ-0XX-p1600-t1-r1 is not close to static(<100 games), the score of them is all bellow the 2 hybrid weights, and to my experiment, the score of LZ-0XX-p1600-t1-r1 will continue to drop during 50 to 100 games process.
So after several test, in my opinion: