aki65 / aki65.github.io Goto Github PK

it would be convenient: if default_model.bin.gz is in the folder with the apk file, then when you start BadukAI it is loaded. probably better at start launching KataGo

I run BadukAI on my Android phone, but it is weaker than KaTrain or Sabaki with KataGo on my laptop when config the same number of visits. So I want to find the default_gtp.cfg on my phone to see what is the cause. But I can't find it.

The "__go4go_branches.sgf.txt" below is too big and BadukAI is too slow...
__baduk_branches.sgf.txt
__baduk.sgf.txt
__go4go.sgf.txt
__go4go_branches.sgf.txt

thanks for your hard and wonderful work. here is a suggestion. many people ，i think, prefer a weight that is much faster to run,not that the stronger the better. after all, most of us is not beyond pro-level,and 6-block or 10-block weight is enough and suitable.
so could you please offer a optimized 6-block or 10-block katago weight?looking forward to your reply.

Thanks for the nice app, could you add support for other sizes(KataGo only)? 9x9 and 13x13

I just want to confirm the findings of Vargo in https://lifein19x19.com/viewtopic.php?p=262439#p262439.
On my phone (OnePlus 8 Pro with Snapdragon 865 and Android 11) katago stopped working on all nets with the optimizations in v0.18.0.
Sorry, for not reporting this earlier :-(

for example BadukAI v14 3d 15b - CS Zero 3d 7-3, BadukAI v15 3d 15b - CS Zero 3d 0-10. so it should be? all settings are the same.
also plays poorly against CS Zero 5k, BadukAI v15 3d 15b - CS Zero 5k 0-1.
BadukAI v15 3d 15b - CS Zero 11k 0-1
plays like a beginner. maybe there is some mistake?

Android notetice : View in Full Screen.
Then baduk crash.
Xiaomi Poco X3 GT - Full phone specifications
dotOS Android 11
Baduk_1.9, full and red.

So I can know whether hardware accelation is worked?

My device has a 768g soc.

From https://katagotraining.org/networks/,I only find the .bin.gz files,but not find any tflite files.

where can I find those files?

thanks.

Just test Baduk on it, only got 20n/s for katago 20b, which is much slower than 768G (80n/s).

Anyone can confirm?

My device is Vivo T1.

It might be just me, but I feel like the kyu bots are weaker in version 0.16.0 than 0.14.0. I use the kyu rank 10 bot with the 6b net. In 16, if you attack a group in the beginning, eventually the bot will give up defending and you can always capture it. In 14, you can never just capture a group that can easily escape, unless the bot is strategically sacrificing the group for territory elsewhere. it will trap you instead. Thanks!

Whenever to evaluate a GO game, the numbers of capture stones of both players must be counted. However, it seems that BadukAI doesn't display the numbers and has no spare space to show.
In my view, the numbers can be placed with "komi" in the setting menu.

I set resign-score=30 and resign-winrate=-0.80, then I play kyu-rank mode from 30 to 1, it seems that KataGoes resign later than the board positions, which are over 30 or -0.80 if you using ponder.

Is the distributed training of KataGo begin? Where are they?
How many blocks do they have, 50b or 60b? Are they stronger than g170-b40c256x2-s5095420928-d1229425124 and have more ELO than 1919?

The Chinese was translated with google?

I think it might be stronger than the current 15b

To get small apk size and use additional optimized network for katago.

能不能增加支持手机gpu运算

could you please add optimized human-style weights? for example，leelamaster GX78,leelamasterGXA4,leelamasterE08,anyone is ok.I could provide you with them,if you need.you can get them here (https://github.com/pangafu/LeelaMasterWeight)also.looking forward to your reply.

Hi, thanks for your amazing app! I am trying to figure out how the optimization works. And I just checked the nodes of the .tflite file. Some of the them seems to be removed from the original nets? Could you possibly give some details about its structures. I just want to write some scripts to make the conversion more convenient. Thanks!

Thanks for your work, that people can use LZ and KG on android devices!

How my W winrate and score-lead be displayed:

W winrate:
97.3%

score-lead:
11.2

Only a half of string could be displayed, the other half disappear.

I'm using Xiaomi Poco X3 GT with MediaTek MT6891Z Dimensity 1100 5G (6 nm).

ROM: https://developer.android.google.cn/topic/generic-system-image/releases aosp_arm64-exp-S2B3.220205.007.A1-8133516-58a8796c

1.13.2 optimized 40b : 10v/s
1.13.0 build in 20b : 20v/s

Can we add support for GPU operations? If not, what are the obstacles?

BadukAI has "v/m in analysis", how could it be?
I find "kata-analyze" (from KataGo/docs/GTPExpension) is hard to control somtimes. The reason of it maybe the output of KataGo and the fetch-pointer (be used for reading the output of KataGo) are asynchronous sometimes.

Obviously the "KG 20b" is not a official network, it's much faster and seems much weaker. Is it 20b128f(b20c128)?

https://lifein19x19.com does not work today, so I am writing here.
0.15 - installation error, on the emulator

I don't know what is the final goal of BadukAI, I also don't know how many programming skills do you have, but these features are useful and convenient:

• read different file directly, for examples .txt .zgf(ZenithGo) .gib(www.eweiqi.com)

• convert sgf into pdf, static image, gif, for examples:
  
  

• create sgf from camera or photo, for example, generate sgf from the "leesedol vs alphago game 4" static image above, recognizing the move numbers on the stones as the order of the stones

• Real Wood Board Theme, from Sabaki:
  

• visualization of ownership predictions, for examples,
  

• analyse all based on playouts or visits, now "analyse all" may skip some moves without giving a histogram if the sec/move is too small or the board position is too complicated.

• accelerate "analyse all" if several continuous board positions are on the same known tree, I don't know if it make sense. Another: If "analyse all" from end to begin, could it pick up speed? Because the whole game tree are known.

• clock mode, for examples, 5min+10sec5, 30min+00, 0+30sec*3

• add local page "How to Use" (the same to https://aki65.github.io/) to app

It's said v1.9.1 75 Elo stronger than v1.8.2 with the same latest neural nets.

katago最近引擎更新，似乎提升很大，什么时候能够应用到badukai

in v1.5.1 it is impossible to connect lz network. then lz can be excluded from the application

I can solve those problems by shrinking the size of chessboard through Sabaki on my laptop. But how to solve those problems on my Android phone?

I hear about it that the GTP command "ownership" could slow down analyze speed a lot, is it true in BadukAI?
How many times does BadukAI calculate ownership, only once after a ownership-query(a user hit "own" button), or every time meanwhile every analyze?

Can players add a new teaching mode when playing with AI? When the lost points exceed the set limit, the AI will automatically delete and let the player choose again.Just like the software KaTrain on the computer.This will help the player to learn the AI idea better.
@aki65

• Analyzer and AI-player treat Black as White and treat White as Black after editing static position sometimes.

  • 
  • 

• Detecter doesn't remove illegal stones automatically.

  • 

Because BadukAI has not many items, I could translate into Chinese in a day.

for example playouts per second?

Acceleration weight is not compatible with snapdragon 888.There will be a wrong message.But it also works, just not that fast,Just like a regular one.Please see the attachment I uploaded for the wrong information.

Hello

I used the 1.13 version.

I'm a starter player and as such, I wanted the AI to be at a low level.

I first tried a 19x19 grid, at 30 kyu-rank with the 20b network. AI is black but even for me, il seems to choose its moves randomly. I can try to surround its stones and it never stops me. I can put it in atari and it never protects its stones, etc. etc. After we reach what seems to be the end of the game, AI still fills the grid, it never passes nor resigns.

30k is low and my game is poor. I tried to let it play against itself 19x19, 30k and 20b network, and the result is the same: it fills the grid and captures stones indefinitely and never passes or resigns. Same if I try with 1 (1k) or -8 (9d) kyu-rank.

I tried several configurations of the AI playing againt itself, also with 9x9 grids, but as soon as I put a kyu-rank, the bot doesn't "work" any more.

I finally tried a 9x9 grid against itself, with no kyu-rank and in this case, it plays correctly and black or white wins after the two passed.

Am I missing something?

Turn off 'winrate on play' then turn on 'analysis all', the 'winrate for W' and 'score-lead for W' will disappear.

Analyzing a whole game of more than 300 moves step by step seems too slow. Could you add KPAE to BadukAI?

The move-winrate histogram may become readless and meaningless when winrate changes in a slit of 90%-100% as the scorediff is over 8. And a game of 300 moves will come to an end at move-30 because the winrate is around 95% from move-31 to move-300. So could you add move-score histogram?

I tested a lot of sgf, it seems that the winning rate of lz engine is always between 12% and 88%.

The kata engine is no problem.

The analyzed results of different GUI are different, i.e. Sabaki can't show winrate graph when reading the analyzed results generated by KaTrain. They have no universal format.
Will BadukAI write&read analyzed results? What the format will be?

It seems that the optimized 20b Kata weight is not supported more than komi 10?
When I set the komi to 20, I get the same winning rate as when I set the komi to 10.

But the ordinary 20b weight does not have this problem.

Is it possible to support Android 4.4 by compiling it with lower version of Android SDK? Thanks!

I use a e-ink device to make my eyes feel comfortable. So can this Go AI APP provide a black And white theme for E-Ink device?

Hi, thanks for your amazing app! I am wondering if there could be some more custom settings in the config. Like (dynamic) PDA (to deal with the handicap games more properly), maxTime (to make the search time more flexible while playing against katago). By the way, how did you convert the orignal network to a smaller one? Can I do this on my own so that I am able to use some other networks? And could this optimized network be applied to the katago engine as well? Thanks! @aki65