In the face of the power of minigo v15 000939 , 40b completely overestimates their winning percentage

[ainixingo]

https://www.youtube.com/watch?v=FKinqyTv4h0&t=505s

However, this is not a single example

[l1t1]

could you share a package of the minigo

[l1t1]

I got it from qq group
https://userscloud.com/wd0tqdqkqvia

D:\>d:\leela-zero-0.16-win64\leelaz.exe -w 939-heron.gz
Using 2 thread(s).
RNG seed: 4878283789657530464
Leela Zero 0.16  Copyright (C) 2017-2018  Gian-Carlo Pascutto and contributors
This program comes with ABSOLUTELY NO WARRANTY.
This is free software, and you are welcome to redistribute it
under certain conditions; see the COPYING file for details.

BLAS Core: Haswell
Detecting residual layers...v2...256 channels...19 blocks.
Initializing OpenCL (autodetecting precision).
Wavefront/Warp size: 32
Max workgroup size: 1024
Max workgroup dimensions: 1024 1024 64
Using OpenCL half precision (at least 5% faster than single).
Setting max tree size to 4077 MiB and cache size to 453 MiB.

Passes: 0            Black (X) Prisoners: 0
Black (X) to move    White (O) Prisoners: 0

   a b c d e f g h j k l m n o p q r s t
19 . . . . . . . . . . . . . . . . . . . 19
18 . . . . . . . . . . . . . . . . . . . 18

[l1t1]

990 is uploaded here
https://userscloud.com/cat842csy8es

[l1t1]

1005 _quantized 2-3
https://userscloud.com/w598ji58xolm

[alreadydone]

I quantized it with your modified script :)

[roy7]

I tossed my own converted 939 up for a test run. Didn't do well, but it does much better in a time parity test (it's a 19b network). I'm using that net for the RoyalMinigo bot on OGS currently.

[sethtroisi]

Andrew and I played a 100 games at 5 seconds per move between minigo v15-939 and 201 if you're interested in more games.
Ringmaster control file, Games on CloudyGo and final report (Spoiler it was a nail biting 50-50)

[sethtroisi]

I also have a patch I'm testing that will add winrate to sgfs so that future ringmaster games will have eval curves

[l1t1]

@roy7 mini939 is much stronger than elfv1 read from the elo?

[barrtgt]

Thanks for sharing the converted weights. Could you make test matches that show up on https://zero.sjeng.org/ always run for at least 400 games?

[gcp]

Networks don't have perfect play or score estimation, nothing new and nothing to fix here.

[barrtgt]

The reason I requested more games for the minigo match was because it is our only independent benchmark that our clients can run to test against and 83 games is not nearly enough to be meaningful.

[l1t1]

if there are some minigo weights stronger than lz of that time in the future, (it is possible because of google's resources)
will we use them to generate self play games to help lz?
just like elf v0 and v1 did some months ago.

[l1t1]

minigo v16 40b is training
https://cloudygo.com/