Best Nets for Lc0 Page

[mwyoung]

With a few simple tweaks this net series is testing 50 to 60 elo better then default settings.

Share your Settings please.

+1

I dont know your hardware or net or setting. But for the new T61000 nets. And other large nets. Change the temperature policy. Default is way too high. This is one of the main parameters to change. Set it from 2.2 to 2.0. This should help a lot.

What this does is narrows the search allowing for a much deeper search in some positions. This cures many of the tactical oversights seen in default setting.

[Robert Flesher]

With a few simple tweaks this net series is testing 50 to 60 elo better then default settings.

Share your Settings please.

+1

I dont know your hardware or net or setting. But for the new T61000 nets. And other large nets. Change the temperature policy. Default is way too high. This is one of the main parameters to change. Set it from 2.2 to 2.0. This should help a lot.

What this does is narrows the search allowing for a much deeper search in some positions. This cures many of the tactical oversights seen in default setting.

Thank-you! I try that out.

[bmp1974]

In the Lc0 network page, what is the "run" 1 or 2 indicate? And why is T50 series network file is smaller in size compared to T40 or T60 series?
Appreciate if someone can throw light on these.

[Hugo]

Hi computerchess enthusiasts

Finally I found a first Lc0 challenger to Stockfish.
It is Leelenstein 12.2.

Code: Select all

   # PLAYER                           :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 Lc0 v0.23.1+git.-leele12.2       :      12     14   155.0     300  51.7      85   25   260   15  86.7
   2 Lc0 v0.22.0-leele12.1            :       1     14   150.5     300  50.2      55   20   261   19  87.0
   3 Lc0 v0.22.0-leele11.1            :       0     14   150.0     300  50.0      50   16   268   16  89.3
   4 Stockfish 031219 64 BMI2-T13     :       0   ----  1061.5    2100  50.5      63  157  1809  134  86.1
   5 Lc0 v0.22.0-sergio1207           :      -2     14   149.0     300  49.7      64   15   268   17  89.3
   6 Lc0 v0.22.0-T40.B4-160           :      -6     15   147.5     300  49.2      74   24   247   29  82.3
   7 Fat Fritz                        :     -13     15   144.5     300  48.2      70   18   253   29  84.3
   8 Lc0 v0.22.0-61653                :     -19     15   142.0     300  47.3     ---   16   252   32  84.0

https://docs.google.com/spreadsheets/d/ ... =609748570

from now on I will change my test hardware, wasnt satisfied about the hyperthreading .

kind regards, C.K.

[mwyoung]

In the Lc0 network page, what is the "run" 1 or 2 indicate? And why is T50 series network file is smaller in size compared to T40 or T60 series?
Appreciate if someone can throw light on these.

The run 1 and 2 just indicate that 2 networks are being trained. Run 1 is T60, and 2 is for T50.

T50 is designed to run on weaker hardware like a CPU or slow GPU and still get good performance only 128 filters.

T40 with a filter size of 256 was the previous top performer on the Lc0 network page.

T60 with more filters 320 is an attempt to surpass the T40 networks. The theory is bigger nets should play stronger as they take longer to saturate. Meaning they can be trained longer with elo gains.

[bmp1974]

In the Lc0 network page, what is the "run" 1 or 2 indicate? And why is T50 series network file is smaller in size compared to T40 or T60 series?
Appreciate if someone can throw light on these.

The run 1 and 2 just indicate that 2 networks are being trained. Run 1 is T60, and 2 is for T50.

T50 is designed to run on weaker hardware like a CPU or slow GPU and still get good performance only 128 filters.

T40 with a filter size of 256 was the previous top performer on the Lc0 network page.

T60 with more filters 320 is an attempt to surpass the T40 networks. The theory is bigger nets should play stronger as they take longer to saturate. Meaning they can be trained longer with elo gains.

Thank you mwyoung

[Robert Flesher]

Sergio 1705 net is out. Strong!!

https://groups.google.com/forum/#!topic ... SbCcfEzdo0

[Hamster]

Very interesting thread and I understand that with short time controls a small net is necessary as the large one would not have enough time for searching "sufficient" nodes. But how many nodes are sufficient? Is there a minimum number of nodes lc0 should be able to search with any net? If not then the largest net would always be the best.

[mwyoung]

Very interesting thread and I understand that with short time controls a small net is necessary as the large one would not have enough time for searching "sufficient" nodes. But how many nodes are sufficient? Is there a minimum number of nodes lc0 should be able to search with any net? If not then the largest net would always be the best.

If you have a fast card, as in any RTX card. I would not mess around with the smaller nets at any time control. On this hardware the big nets are still best IMO.

And if running a slow card, or CPU. That would be the only time I would run use the smaller nets.

Currently the best nets can be found here. https://github.com/jhorthos/lczero-trai ... a-Training

[Hamster]

I have a cheap Radeon RX550 and I am using lc0 v0.26.2 (with Nibbler).

With the default net (703810.pb.gz) I get in about 1 minute: 154k nodes (2400N/s)
With the J92-145 net I get in 1 minute: 7.7k nodes (120N/s)

Which one should I use (assuming the same amount of time will be spent)?

Default net


J92-145