KataGo V1.3

[lightvector]

KataGo 1.3 released! Lots of new features, and neural nets from the new run!

All the details here:
https://github.com/lightvector/KataGo/releases/

There's always some chance I overlooked some major bug or problem in my testing, so apologies in advance if you find any such thing - please let me know and I will fix it over the next several days.

Edit: And a quick followup release is out, v1.3.1, along with a stronger net from the current ongoing run! (updated the link above)

Edit: And now v1.3.2 is out! Same link above, but better nets, better OpenCL performance, more bugfixes and fixes to missing features.

Edit: And now v1.3.3 is out, with some early versions of some new big neural nets that are training, better 20-block nets, and a variety of interface improvements and minor feature additions.

Edit: The latest version is now v1.3.5, and also there are some nets more recent than that with significant strength improvements.

[Knotwilg]

KataGo 1.3 released! Lots of new features, and neural nets from the new run!

All the details here:
https://github.com/lightvector/KataGo/releases/tag/v1.3

There's always some chance I overlooked some major bug or problem in my testing, so apologies in advance if you find any such thing - please let me know and I will fix it over the next several days.

Hi, I wanted to replace my Kata in Lizzie with the new version.

1) I replaced the network by one of the attached networks
2) I replaced the KataGo folder with an unpacked windows version of KataGo as published in your link.

Engine keeps loading. What did I do wrong?

Thanks

[ez4u]

First you should run your new Katago 1.3 from a command window. That way you can see any error messages from Katago itself. I was using the cuda version and it could not find cudnn64_7.dll so I copied it to the directory. I ran the benchmark command and then the gtp command. After you can get a reply to "genmove black", you are ready for Lizzie.

Did you copy the new config file to the Lizzie directory? It is unclear in your post. I set up version 1.3 as a new engine. Needless to say, I made a mistake typing in the command line into lizzie. I run Lizzie from a command window and that allows me to see any error messages from Java, in this case that it could not find my file.

After these hiccups, 1.3 runs fine for me.

[lightvector]

@Knotwilg - please see also the section about Upgrading on the releases page above: https://github.com/lightvector/KataGo/releases/tag/v1.3

It is highly likely you will also want to discard your old config, except to copy any settings you specifically changed into the new config, otherwise you will be missing a few features and improvements. If you didn't ever edit the old config that Lizzie came with either, then probably also KataGo was running with not as good performance on your computer as it could have before.

To better adjust settings for good performance, and to debug why KataGo isn't working in the first place, both are easier to solve if you run the benchmark command directly in a console first rather than trying to immediately use KataGo with a GUI.

[And]

when to expect a version or option cpu-only? and versions with support up to 25x25?

[Gomoto]

Both networks up and running with Lizzie 0.7.2
I like fast loading of the 10 block network very much

Thank you lightvector!


(And no more coil whine with V1.3 and the new config by the way )

[xela]

Leela Zero has recently been hinting that its strength is approaching a plateau. If I understand correctly, KataGo has had a fraction of the training (in terms of CPU-hours or number of games) of LZ. Is it possible to get any hints (from learning rates or other metrics) whether KataGo is also approaching the ceiling of its strength, or whether it could go significantly beyond LZ given enough training?

[lightvector]

KataGo should not plateau for a long time, or at least I will run out of my compute budget before that happens.

The current 20-block network probably still has 200 Elo (hopefully more than that!) to grow beyond the released version, which might actually put it not too far from visit-parity with Leela Zero's 40-block (and therefore quite a bit stronger at equal compute time).

KataGo has a 30 block or a 40 block brewing which theoretically should have quite a bit more capacity than that, which we will switch to at some point.

[lightvector]

Also, I suspect Leela Zero is not plateauing, or if they genuinely are, then minimal changes such as to the training window, game variety, gating, etc, should "unplateau" them. I'd be a little surprised if a 40-block network cannot improve further, if KataGo's merely 20 block network is going to be approaching its strength.

Except for the network *actually* becoming "saturated" and finally capping out, the issue is that Elo gain is roughly proportional to log(selfplay compute spent). Log is a very, very slow growing function, so it can feel like you're on a plateau even if you're not. But different enhancements can translate the curve in logspace so that you are N times faster all the way through, which is hopefully how KataGo might pass the point where Leela Zero *feels like* they've plateaued, we'll see.

Or maybe you feel that log grows so slowly that once you are far enough into the tail of the log you call it a plateau anyways. That's also okay.

[Bill Spight]

Also, I suspect Leela Zero is not plateauing, or if they genuinely are, then minimal changes such as to the training window, game variety, gating, etc, should "unplateau" them. I'd be a little surprised if a 40-block network cannot improve further, if KataGo's merely 20 block network is going to be approaching its strength.

There is the possibility that LZ is near the top of a hill. (Not that I have any idea, but it is a possibility.) If so, KataGo's different architecture may be causing it to climb a different hill.

[xela]

OK, so not technically a plateau (if we believe the assumptions of the Elo rating system). But perhaps near enough for most practical purposes. Thanks for the explanations, this makes a lot of sense.

[jann]

I doubt LZ is near a real peak yet, the same slowdown happened with 15 blocks some time ago, yet the same 15 blocks networks keep getting stronger when fed strong 40b data. Any small change can restart improvement (for a while), it's just that there were no changes so far.

[Limeztone]

lightvector, I'm enjoying your KataGo tremendously.

I use it on my iPad for playing and it works great!

I also experiment a little with it under Sabaki on my PC.
But when I let KataGo play it self with different settings I have no way to see what the instances of KataGo thinks of the position or what the playouts/visits was for a move.

Would it be hard for you to implement the command that Sabaki uses to show that for LZ?
I think the name of the (Sabaki) gtp-command is: genmove_analyze <color> <interval>
https://github.com/SabakiHQ/Sabaki/blob ... gration.md
and will be used by Sabaki if added to the KataGo's GTP. The <interval> argument could be ignored

I don't need the full move list that LZ displays, but the following would be very nice:
- Win% for the move selected
- number of playouts used
- number of visits used
- actual time used

It would be so nice if it was possible to get that information from KataGo

Thanks in advance
Regards,
Limez

EDIT: After scrutinising the .cfg file I tried the sdtErr stuff and it kind of got me what I wanted.
I still think implementing the genmove_analyze command would be very attractive to me.

[lightvector]

v1.3.1 is released now with some bugfixes, as a quick followup to v1.3. And a new stronger neural net!

https://github.com/lightvector/KataGo/r ... tag/v1.3.1

With v1.3.1 you can (and probably should) also now delete a big chunk of your GTP config - defaults are built into KataGo directly and you can just use those directly unless you specifically want to specify some of these parameters to be nonstandard values, to play with KataGo's behavior. The advantage of this is that if parameters improve in future releases (including actually one such change in this release itself), you pick up those changes, instead of your config being hardcoded to old less-good values.

@Limeztone - new version implements lz-genmove_analyze and kata-genmove_analyze, as requested. A quick test with Sabaki indicates that Sabaki seems to be using it now. Enjoy!

[And]

how to set up KataGo in Sabaki on a rectangular board? after each move the warning "GTP engines don't support non-square boards"