That is intended. Before, the buttons were quadratic. And on a phone the top row is much higher (in relation to screen height) than on your tablet, otherwise the text would not be readable. Therefore the buttons are also much wider on a phone,so that they take about half of the screen width altogether. On small phones, this means that there is not enough space for the status text which can be quite long (KG is pondering). So I made the buttons slimmer (less width than height).



The plan was, that the "own" button is as wide as the move buttons, so that there is exactly one move button less on the screen. On my phone it exactly turns out that way, but it seems that the scaling is different on your tablet. I will see if I can find the reason for that and fix it.

and maybe make the buttons of the first and second row from the bottom the same size? then you can move the "own" and "pass" buttons to an empty space. some of the buttons are very small, while some are very large

Wow i cant believe how fast you make the optimized network. Thatis really amazing.

yes, some buttons are too small.
in connection with the end of the training LZ can replace the network with the latter?
the green line on the winrate graph is in stark contrast, perhaps blue or purple would be better?

How many playouts per second is the 40block engine getting?

_________________
- Brady
Want to see videos of low-dan mistakes and what to learn from them? Brady's Blunders

My mistake, I miswrote, I meant how fast he released the net. In terms of playouts, its probably faster than the non-optimized 20b net. On Snapdragon 625, 4gb ram, I get maybe 2 playouts/second, somewhere between 1-2. Which is really nice for these devices

On my Oneplus 7t from 2019 (Snapdragon 855+) I get around 40 playouts per second.

9x9 game, H2, komi 7.5

B: CrazyStoneDL 5d
W: BadukAI, 10s/move, Samsung s20+, network : b40c256-s7087...

Result : W+R

BadukAI wins more than 50% of these games against CSDL (5d, H2, komi 7.5) but usually loses with H2, komi 0.

What a wonderful program !

I released a new optimized version of the s680 network in https://github.com/aki65/aki65.github.io/releases/tag/v1.8.2.

The main difference from the previous version is that it runs much faster on the new snapdragon 888 chipset, so if you use such a device, you should definitely download it. On other devices the speed increase is minimal.

I made a new version (1.9) available on play store https://play.google.com/store/apps/details?id=net.kir.baduk_ai and github https://github.com/aki65/aki65.github.io/releases/tag/v1.9.0 with the following changes:

All built-in networks are optimized in a new way (which also works on snapdragon 888).
The built-in LZ 40b network is updated to final LZ 287.
You can now side-load optimized LZ networks (e.g. the "LeelaMaster GX78" included in this release). To do so, save LeelaMaster_GX78.tflite and LeelaMaster_GX78_tail.tflite in the same directory anywhere on your device. Then tap on the first field in the settings and select LeelaMaster_GX78.tflite in the file dialog (NOT the other one).
Analysis results are now persistent, i.e. if you analyse a whole game and then navigate through the moves, you will always see all the moves and variations analyzed before. Moreover, the analysis results are saved in the sgf on "save sgf" and reloaded on "load sgf".
The sgf reader is more tolerant, so that you should be able to read more sgf files from other go apps or sites.
Some small layout improvements were added.

PlayStore says this app is not available in my country (Russia).
Is this a bug or a feature?

Is there an updated acceleration weight? Now the strongest is
kata1-b40c256-s8003455744-d1941206379

It's a bug. I corrected it. Hope it works for you now.

Thank you!
I've already installed the app via guthub
Works fine!

I published an optimized version of kata1-b40c256-s8096761856-d1963946463 at https://github.com/aki65/aki65.github.io/releases/tag/v1.9.2

Aha, thank you very very much!!

What's the version of katago in baduk_ai? Is it 1.7.0?
Now the official katago version has been updated to 1.8.2!

BadukAI uses a revision of katago that is just a few commits before release 1.8.0. So it still issues "1.7.0" as its version number, but the functionality is pretty much identical to 1.8.0.

can you provided an optimized 15b or 60b network? I want to test it

The latest 15b network is about 1000 Elo weaker than the latest 20b network, so even at time parity 20b is much stronger, so I don't think that it makes sense to invest time on 15b.
60b would be very interesting, but I need the raw TF weights (the second download column on kata1 webpage) and these were not yet published for any 60b network. As soon as they are available, I will try on that one.

KataGo 1.9.0 has released!

It say KataGo might be somewhere around 75 Elo stronger than v1.8.2 with the same latest neural nets. It's a very big improvement! I hope BadukAi can be used it soon.

And the weight s908 is much better than s809, I hope it can be updated as well. ^-^

Thank you very much!