Leela network questions

[Guenther]

I can see there are several networks for download, but Leela uses only one of them at a time.
What is the difference between the networks? Do they impact playing strength? Or playing style? Which is the one to use?

Another question: does Leela learn from playing in a tournament against various opponents?

1. Of course they impact playing strength, because all 'knowledge' is in the NN file

2. Since NN ID123 the network size was increased to add more blocks+filters (from 6*64 to 10*128) and there should be some bigger jump in strength coming soon.
There was also a change to affect variety of opening moves in selfplay, which should also result in more strength (in real play) to previously 'neglected' opening trees. (I guess this needs quite some time)

3. You should always use the latest one (which did not fail = green),
even if it appears to have less selfplay rating, because this does not correlate to strength in games vs. other opponents.

4. The NN learns only in training mode

5. If you want to start testing it now it is suggested to do some 'fulltuning' of your gpu to get the most out of it. This will increase the nps of LCZero,
otherwise it quickly creates a working tuning file in the first run which is not optimized, but works with your gpu.

e.g. (might take a while - this produces a file 'leelaz_opencl_tuning')

Code: Select all

lczero.exe --tune-only --full-tuner -w ID123

[Guenther]

Thanks Günther.

*added another point, while I was editing my post

[Werewolf]

I can see there are several networks for download, but Leela uses only one of them at a time.
What is the difference between the networks? Do they impact playing strength? Or playing style? Which is the one to use?

Another question: does Leela learn from playing in a tournament against various opponents?

1. Of course they impact playing strength, because all 'knowledge' is in the NN file

2. Since NN ID123 the network size was increased to add more blocks+filters (from 6*64 to 10*128) and there should be some bigger jump in strength coming soon.
There was also a change to affect variety of opening moves in selfplay, which should also result in more strength (in real play) to previously 'neglected' opening trees. (I guess this needs quite some time)

3. You should always use the latest one (which did not fail = green),
even if it appears to have less selfplay rating, because this does not correlate to strength in games vs. other opponents.

4. The NN learns only in training mode

5. If you want to start testing it now it is suggested to do some 'fulltuning' of your gpu to get the most out of it. This will increase the nps of LCZero,
otherwise it quickly creates a working tuning file in the first run which is not optimized, but works with your gpu.

e.g. (might take a while - this produces a file 'leelaz_opencl_tuning')

Code: Select all

lczero.exe --tune-only --full-tuner -w ID123

Thanks, very helpful post

[MonteCarlo]

Just to be clear, the change to get more variety in openings (the temperature decay) is only for the match games used for gating (determining whether the new network passes or not); it is not used in self-play games.

There's already a high variety of openings in self-play because a temperature of 1 is used for the whole game.

In the matches, where temperature wasn't used, they were getting crazy things like half the games starting from the same position after 8 moves, which they were worried was making the match results less informative.

That's what the change is meant to address.

[carldaman]

5. If you want to start testing it now it is suggested to do some 'fulltuning' of your gpu to get the most out of it. This will increase the nps of LCZero,
otherwise it quickly creates a working tuning file in the first run which is not optimized, but works with your gpu.

e.g. (might take a while - this produces a file 'leelaz_opencl_tuning')

Code: Select all

lczero.exe --tune-only --full-tuner -w ID123

Hi Guenther,

Is full-tuning to be done for cpu's as well, or just gpu's?

Thanks,
CL

[Jhoravi]

In the matches, where temperature wasn't used, they were getting crazy things like half the games starting from the same position after 8 moves, which they were worried was making the match results less informative.

That's probably the reason why Google didn't publish all 100 games between Stockfish and Alpha0. Since no book to vary SF opening, it keeps on repeating the same loosing variation in the Queens Indian Defense as shown in their selected 10 game sample. It also explains why too many looses as black(27) than as white(3)

[MonteCarlo]

Full tuning is only for the GPU version.

Cheers!

[carldaman]

Full tuning is only for the GPU version.

Cheers!

OK, good to know, thanks!