Latest 11258 distilled network for CPU released: 112x9-se

[Dietrich Kappe]

Can be found here: https://github.com/dkappe/leela-chess-weights/wiki/Distilled-Networks

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)

   1 ethereal                :    3360     51   290.5     320  90.8     100  267   47    6  14.7

   2 ID11258-112x9-se        :    3098     58    67.0     104  64.4      91   49   36   19  34.6

   3 crafty25.2              :    3051     32   230.5     363  63.5      69  189   83   91  22.9

   4 ID36092                 :    3036     49    88.0     146  60.3      59   58   60   28  41.1

   5 ID11258-96x8-se-5       :    3029     34   169.0     296  57.1      55  124   90   82  30.4

   6 ID35975                 :    3026     39   145.5     252  57.7      66  103   85   64  33.7

   7 ID11248-128x10-se       :    3016     31   220.5     380  58.0      79  167  107  106  28.2

   8 ID35689                 :    2995     44   103.5     175  59.1      52   70   67   38  38.3

   9 cheng4                  :    2993     65    37.0      73  50.7      85   27   20   26  27.4

  10 amoeba                  :    2949     50    68.0     144  47.2      61   45   46   53  31.9

  11 scs-64x8-run1-550000    :    2938     60    42.5      96  44.3     100   31   23   42  24.0

  12 ID11248-256x12-se       :    2808     54    41.0     126  32.5      62   22   38   66  30.2

  13 winter                  :    2796     54    34.5     129  26.7     100   21   27   81  20.9

  14 crafty19.18             :    2693     38    79.0     363  21.8      66   53   52  258  14.3

  15 ID11258                 :    2679     64    26.0     124  21.0      95   15   22   87  17.7

  16 ID11258-16x2-se-3       :    2592     73    21.5     126  17.1      58   11   21   94  16.7

  17 ID11258-16x2-se-4       :    2583     66    24.0     159  15.1     ---   12   24  123  15.1

White advantage = 45.21 +/- 8.08

Draw rate (equal opponents) = 36.60 % +/- 1.51

[Alexey Eromenko]

How is a distilled (reduced) Network stronger than original? Was it re-trained ?

[Dietrich Kappe]

When you train a network, you feed it the history and position info and the value (game outcome) and policy from the move made (supervised learning) or the policy values from the 800 node search.

In distilling, you use regular training data for history and position info, but value and policy comes from the teacher network. And yes, you could distill to a bigger network.

The distilled network here is stronger on cpu, because the original network runs too slow. If you were to run them on gpu, the original network would be stronger.

[sherief...@gmail.com]

but lc0 running on CPU only is alot weaker than GPU

Op zondag 13 januari 2019 21:39:42 UTC+1 schreef Dietrich Kappe:

[Dietrich Kappe]

Yes, but some folks don’t have a gpu or only a very weak onboard one.

I want to give them something with leela’s style and flavor, which can be run on cpu. Adding Ender will hopefully push it over ethereal on a single core.

[josé luis]

Thanks Dietrich..you are great

[Jon Mike]

Thank you, thank you, thank you! 

:)

On Sunday, January 13, 2019 at 4:12:30 PM UTC-6, josé luis wrote:

Thanks Dietrich..you are great

[garrykli...@gmail.com]

wheere instruction how make it work with fritz gui

[Dietrich Kappe]

You will need a recent lc0 executable to run the nets. You can find v0.20.1 here: https://github.com/LeelaChessZero/lc0/releases/tag/v0.20.1

If you are running cpu, you want the blas flavor.

As far as setting it up, it’s a uci engine. Point your GUI at it. The three uci options worth changing are:

option name WeightsFile type string default <autodiscover>

You’ll want to point it at wherever you placed the .pb.gz weights file.

option name DirichletNoise type check default false

You’ll want to set this to true to get a little variety in play.

option name SyzygyPath type string default

You want a semicolon separated list of dirs where you keep your endgame tablebases.

Best of luck.

[FWCC1]

have you added Ender?

[Dietrich Kappe]

I’m distilling a 200x20 net at the moment. 128x10 is already released for Ender. https://github.com/dkappe/leela-chess-weights/wiki/Endgame-Net

[Jon Mike]

Here's a link to all Dietrich's weights for Lc0

On Sunday, January 13, 2019 at 2:39:42 PM UTC-6, Dietrich Kappe wrote:

...

[Jonathan Rosenthal]

 Awesome! Seems that under your conditions you are slowly converging to an optimal size. Does "ID11258-96x8-se-5" signify that it was your fifth attempt with that network size or does the 5 at the end signify something else?

I do wonder how the relationship between machine speed, number of filters and number of layers works precisely. I also assume search on fast TC on fast machine = search at some long TC on some slow hardware, which would imply that depending on TC a different network may be optimal. I could see lots of other uses for your networks as well, eg: if you think about AB style search you could consider a preliminary search with a small network and then switch to the more expensive bigger net if you need a more precise answer.

PS: Awesome to see that you included Winter in your list as well :D