KaTrain: AI analysis/playing/teaching tool based on KataGo

Thoughts about the older version:
I did play some 19x19 games with the calibrated rank opponent set at 1d.
It felt unnatural. A strange mixture of pro strategy and rookie mistakes. In the opening it makes a lot of tenukis and probes. In many games it first crushes me and then blatantly lets me win.
So then I set it at 3d and most of the weirdness was fixed although a 3d opponent is way too strong for me.
I tried the teaching mode. It forces me to play a very territorial style and give the opponent a huge moyo (no surprise, thats how strong bots roll). And then the game stops because there are 1000 possibilities to invade and I can’t find the only correct KataGoTM magic sequence. Even if I give myself infinite undos it gets tiring to click through every possible empty point. So in this situation it would indeed be super helpful to have it show me options A B C D to pick from.

Edit: by forces I mean encourages, the auto-undo triggers easily when I choose less territorial moves

5 Likes

I’ve found that the teaching mode can at some points allow you to dig yourself into a hole only katago can dig itself out of, but I disagree that it forces you to. The default settings can lead to all sorts of games.
As for calibrated rank, here is how it works, and you’re right in that it’s far away from mimicing human play.

3 Likes

The 1.5.1 pre-release has been out for a while now, but not many people have gotten back to me on the new AI options.
If you can give them a spin and leave your thoughts here, that’d be great. :slight_smile:

1 Like

v1.6 was just released with the new AI option, a katago upgrade, and a whole bunch of smaller features
My favourite one is probably ‘next mistake’ which makes for nice and quick reviewing.

1 Like

Thanks for creating this wonderful tool! :clap:
I switched from Lizzie and have been enjoying using it for reviewing.

An idea for improvement:
How do I cancel/interrupt/pause the full game AI analysis or set it not to start by default? Would be nice to have such feature if it doesn’t exist atm. Some times when I open a game record, I dont necessarily want the full game analysis, but ie. just to go over some specific moves with AI.

GPU intensive apps like KataGo tend to slow down my window manager so I would like to have some control when to start it, ie. when AFK.

1 Like

The tool sort of relies on analysis being there (or pending) for every node.
I recommend just setting fast visits in general settings to something low, like 5 if you want really fast loads (and check the box in the load dialog).

v1.7 was just released with a bunch of new analysis options, including tsumego solving, and much smoother handling of sgfs.

1 Like

Why?
I just need analysis for current move.

I opened 40 block model on weak small laptop. Even 1 visit for every of 200 moves of game is long time total. I review fuseki only - I need STOP button - so it don’t eats battery energy while thinking about moves I don’t care about.

Update:
Oh, there is delete node button - but “just stop thinking” button would be more elegant.

Nice added features, thank you.

Did Katrain open the menu before if you switched the window with Alt + Tab? It does now, anyway. I use it often, for example if I replay a game from a Youtube video. Using Alt as a shortcut isn’t a good idea I think.

Interestingly, the menu also appears when I switch with Win+Tab (left Win), although Win does not seem to be used in Katrain.

Why are you using the 40 block network? Basically only top-end machines can use that effectively.
Less blocks means each visit is less accurate, but significantly faster, leading to better results faster. Keep in mind kata usually requires around ten thousand visits to get trustworthy results.

1 Like

I like neural networks - its not soulless brute-force or monte-carlo tree search
Its truly made artificial intuition
I can’t beat 1 visit 40 block even with some stones handicap - so its more than enough
I guess patterns from such game may be more useful - they are less random than when ai uses many playouts - they are ignoring what normal ai takes in consideration - so its more easy to replicate to human

Any PC takes forever to get the big network up and running. And then the UI becomes unresponsive. In an equal CPU time match the smaller network tends to lose to a bigger network but that should only matter in a computer go tournament. So for an interactive user experience the smaller networks are much better. Even 15b is too strong for most users. A 5d teacher should be enough for a kuy player, no?

It used to be shift, but that interfered with other things. I should probably switch it to keyup instead of keydown.

I think, Alt or Shift would both be ok if you only react on key up and if no other key is pressed. This is also how the menu in Firefox works, for example.

V1.7.1 released with distributed training models and a whole bunch of bug fixes.

1.7.2 released with minor fixes, a new theme, tutorial, and an all-in-one OSX package.
By the way, if anyone is interested in creating themes, I am happy to help package them. I know it’s still a bit technical.

Oh man! Just discovered this. So excited to try it out tomorrow. Sounds like such a great idea. Bravo!

1 Like

v1.8 was just released with support for helping distributed training, and a feature for generating uneven positions to play from and practice your mid- and endgame skills.

5 Likes

v1.9 released, with a little competition on theming here

1 Like

v1.11 released

3 Likes