Life In 19x19

Posted: **Fri Jan 14, 2022 3:32 am**

akigo wrote:Unfortunately I don't own a phone with SD 845, so I can't analyze the problem. Thus I can only hint at a few potential reasons for this:

1) Model structure
The code in BadukAI for running KG models hasn't changed for a long time, but the models themselves have (including structural changes). I wouldn't expect these to break compatibility with SD 845, but you never know. To check that, you could load the first published 40b model (from https://github.com/aki65/aki65.github.i ... tag/v1.4.1) and check if that one works.

2) Permissions
There are a lot of different file access permission models on different android versions, so accessing external model files can always be a problem. To check that, you could load the internal 20b network (it's also optimized) and check if that one works.

3) Operating system
Did you receive a system update from your phone's manufacturer recently ? I'm asking because BadukAI needs to access the neural processing unit (NPU) to run the optimized models. The interface, that it uses on SD 845 (and many other chipsets) to do so, was recently discovered to have some security flaws in rare circumstances. Some vendors took the easy way out: Instead of fixing the problem, they blocked the interface completely. I had this on one of my own phones some time ago. Very annoying, but nothing to be done about it.
If that also happened on your phone, it would explain the crash.

Thanks for your throughout answer!

1) I tried the old weights (1.4.1) and they didn't work either so that probably excludes the first option?

2) I believe it's not a permission problem since the app can load normal weights that have been downloaded from katagotraining.org and saved to ext media. Also internal 20b network doesn't work but the internal 10b weights load fine.

3) I have updated my phone's FW/OS to the latest so maybe it's the third reason.

Posted: **Sat Jan 22, 2022 8:30 am**

I made a new version (1.13) available on play store https://play.google.com/store/apps/deta ... r.baduk_ai and github https://github.com/aki65/aki65.github.i ... ag/v1.13.0 with the following changes:

1) Option to play "One color go" (more details at https://aki65.github.io, search for "One color go")

2) Detect board and stones from images not only for full boards, but also for parts of a board, like tsumego etc (more details at https://aki65.github.io, search for "load sgf"). This now also works for board sizes 9x9 and 13x13 (just set this board size before tapping "load sgf").

3) The button "Load last game" now really loads the full game (with all moves), not only the final position. Loading only the last position is still possible with the new button "Load last board position".

Posted: **Sun Jan 23, 2022 8:09 am**

akigo
thanks for the update and thanks again for a great program!!!

Posted: **Sun Jan 23, 2022 6:16 pm**

Wow! The recognition effect is very good, but can AI automatically recognize rows and columns? It's a little troublesome to count the board lines every time

Posted: **Mon Jan 24, 2022 3:14 pm**

Automatic recognition of the number of rows and columns would indeed be very handy, but it can't be done with the current algorithm (which has to know what to look for). So a completely new approach would be needed to achieve that goal. I'm not sure if I can find the time for doing that and if it will be successfull (To my knowledge there isn't any app or program which can do this in a fast and reliable way, so it's certainly not easy).

Posted: **Sun Jan 30, 2022 5:55 am**

Is there an updated acceleration weight? Now the strongest is kata1-b40c256-s11101799168-d2715431527

Posted: **Tue Feb 01, 2022 2:14 pm**

lovego wrote:Is there an updated acceleration weight? Now the strongest is kata1-b40c256-s11101799168-d2715431527

I published an accelerated version here: https://github.com/aki65/aki65.github.i ... ag/v1.13.1

Posted: **Thu Feb 17, 2022 2:57 pm**

I published a new accelerated network for Snapdragon 888 devices on android 12. For details and download go here: https://github.com/aki65/aki65.github.i ... ag/v1.13.2

Posted: **Fri Feb 18, 2022 3:45 pm**

akigo wrote:I published a new accelerated network for Snapdragon 888 devices on android 12. For details and download go here: https://github.com/aki65/aki65.github.i ... ag/v1.13.2

I have a snapdragon 636 e ink device (onyx boox note air 1) and the speed of this new optimized net is incredible! Is it as strong as the non-optimized net at equal playouts? Thanks for all your hard work!

Set to 5 playouts with numSearchThreads set to 1 (<5 sec per move on snapdragon 636)
Kata 13.2 1:0 vs Golaxy Elephant 1 star Professional level bot B+R
Wow!
https://ibb.co/NtcWvrf

Posted: **Sun Feb 20, 2022 9:21 am**

go4thewin wrote: Is it as strong as the non-optimized net at equal playouts?

I can't answer that question definitively, as that would require a great many test matches and I don't have the resources to run them. But what I can tell is the following:

I run 5-10 test matches against the original network with each optimized network that I publish (just to make sure the optimization hasn't produced garbage due to some processing error). Summing up the results over all the networks I made in the last 15 months, I get the following:
- With 300 playouts each, the original network is a little ahead
- With 350 playouts for the optimized network and 300 for the original one, the optimized network is slightly ahead
Although these results don't prove anything, they suggest that the optimized networks are slightly weaker than their original counterparts, but the strength difference is very small.

Posted: **Tue Feb 22, 2022 4:00 pm**

I know its a lot to ask, but would it be possible to run 5-10 matches of the latest snapdragon optimized network vs the last non distributed training 40b release at 300 playouts each? It is so impressive what you have done. Probably the slight difference is just from the parrallel processing?

Posted: **Tue Feb 22, 2022 9:13 pm**

go4thewin wrote:I know its a lot to ask, but would it be possible to run 5-10 matches of the latest snapdragon optimized network vs the last non distributed training 40b release at 300 playouts each? It is so impressive what you have done. Probably the slight difference is just from the parrallel processing?

the difference in strength is small, it will take a lot of games

Posted: **Wed Feb 23, 2022 3:13 am**

I made a few games, the optimized network is unexpectedly weak

viewtopic.php?p=271178#p271178
EDIT Turned out to be the clock

Posted: **Wed Feb 23, 2022 8:27 pm**

Thanks And for the attempt, I really appreciate the effort! Of course the optimized version may be a little weaker than the same net, but i am very curious how it would compare to the older non-distributed training katago. You gave me an idea, maybe i will try two games at 1 playout. My guess is the optimized version is stronger.

Posted: **Fri Feb 25, 2022 1:13 pm**

go4thewin wrote:I know its a lot to ask, but would it be possible to run 5-10 matches of the latest snapdragon optimized network vs the last non distributed training 40b release at 300 playouts each? It is so impressive what you have done. Probably the slight difference is just from the parrallel processing?

I ran 6 matches (with alternating colors) as you described, and the result was 4-2 in favour of the optimized network.

Life In 19x19

New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"

Re: New android app "BadukAI"