LCZero Network question

[Eduard]

I downloaded the network 36089. It only has 11 MB. That's not much more than destilled networks, running well on CPU only! Question: Why is the network so small? (Normally these networks have around 40 MB)

Btw:
CCRL Rating list: Lc0 w36089 1 CPU 3022 Elo!

Not bad.

[Guenther]

I downloaded the network 36089. It only has 11 MB. That's not much more than destilled networks, running well on CPU only! Question: Why is the network so small? (Normally these networks have around 40 MB)

Btw:
CCRL Rating list: Lc0 w36089 1 CPU 3022 Elo!

Not bad.

The short-lived NN35 series were just 128*10 blocks and filters.
(*18-12-10 †19-01-07 - last one 36092)

http://lczero.org/networks/

[Eduard]

Thanks, that's interesting. On only CPU running show my first tests in analysis mode that these networks (36088 and 36089) could be better than the distilled networks 11258.

[Eduard]

Here for example (but I can post more). After 5 minutes thinking:

[d]r5k1/1b1Nbpp1/np2rn1p/1Bpp4/8/2N1P1B1/PP3PPP/R2R2K1 b - - 0 1

Analysis by Lc0 v0.20.2-rc1 11258 (Distilled) 112x9:

18...c4 19.b3 Sxd7 20.Lxd7 Lf6 21.Tac1 Sb4 22.bxc4 dxc4 23.Lxe6 fxe6 24.Ld6 Sd3 25.Tc2 Lxc3 26.Txc3 Txa2 27.Txc4 Sxf2 28.Tf1 Se4
+/- (0.77) Tiefe: 10/24 00:04:46 47kN
18...c4 19.b3 Sxd7 20.Lxd7 Lf6 21.Tac1 Sb4 22.bxc4 dxc4 23.Lxe6 fxe6 24.Ld6 Sd3 25.Tc2 Lxc3 26.Txc3 Txa2 27.Txc4 Sxf2 28.Tf1 Se4
+/- (0.77) Tiefe: 10/24 00:04:51 48kN
18...c4 19.b3 Sxd7 20.Lxd7 Lf6 21.Tac1 Sb4 22.bxc4 dxc4 23.Lxe6 fxe6 24.Ld6 Sd3 25.Tc2 Lxc3 26.Txc3 Txa2 27.Txc4 Sxf2 28.Tf1 Se4
+/- (0.77) Tiefe: 10/24 00:04:56 49kN

18...c4? is bad, it loses after 19. e4!.

Analysis by Lc0 v0.21.0-rc1 w36088:

18...Sh5 19.Sxd5 Lxd5 20.Txd5 Sxg3 21.Lxa6 Txa6 22.hxg3 f6 23.a3 Ta7 24.Tad1 Kf7 25.Kf1 Tc6 26.Ke2 c4 27.Sb8 Tc8 28.Sd7
+/= (0.59) Tiefe: 13/32 00:04:53 30kN
18...Sh5 19.Sxd5 Lxd5 20.Txd5 Sxg3 21.Lxa6 Txa6 22.hxg3 f6 23.a3 Ta7 24.Tad1 Kf7 25.Kf1 Tc6 26.Ke2 c4 27.Sb8 Tc8 28.Sd7
+/= (0.57) Tiefe: 13/32 00:04:58 31kN
18...Sh5 19.Sxd5 Lxd5 20.Txd5 Sxg3 21.Lxa6 Txa6 22.hxg3 f6 23.a3 Ta7 24.Tad1 Kf7 25.Kf1 Tc6 26.Ke2 c4 27.Sb8 Tc8 28.Sd7
+/= (0.56) Tiefe: 13/32 00:05:03 31kN

18...Nh5 is here one of the best moves.

But I am surprised about the depth. Distilled network 11258 after 5 minutes only in depth 10/24, w36088 in depth 13/32! How can that be?

[Eduard]

Can someone explain me what the difference is between the 50xxx networks with 128x10 (eg NN 50052) and the 30xxx networks 128x10 (eg NN 36089)?

[yanquis1972]

i think 35xx+ were precursor training for t40. it was fully trained (all 4 LR drops) & in some way used to initialize T40 training. it should be much better on CPU than 256x20 nets since it's both mature & fast.

t50 i don't know anything about; it's got some new parameters & uses 10000 visits instead of the prescribed 800. i assume it's experimental in preparation for t60 (but T30 was an experimental run as well). it's very young & several hundred elo below T40 atm, even at 1'+1s TC.

[brianr]

But I am surprised about the depth. Distilled network 11258 after 5 minutes only in depth 10/24, w36088 in depth 13/32! How can that be?

Depth for Leela is an artificial estimate and not really like a/b engines.

[Eduard]

I understand, thank you!