Life In 19x19

KaTrain simple style(2d) - CS Zero 9d 1:1 (W+, W+)

I made a new version (0.18) with the following changes:

1) If (and only if) you have loaded an sgf with only set stones (that is without any moves), you can now delete stones by placing a move on an already existing stone (useful to edit a position to analyze).

2) The built-in 20b KG network in the 64 bit version is now optimized for android. On CPU it should run about twice as fast as before. Furthermore it will try to use the NPU if possible. The effect depends very much on your device. On my handset with Snapdragon 855+ it is about 6 times faster than the previous version.

The second change is quite experimental as it is very HW-dependent and I could test it only on my device. If you experience problems, please tell me.

Amazing speed increase. With the new optimized net on my snapdragon 625, with the default settings it plays at 9d bot strength, even if you reduce it to just 1 thread!

does it matter in what format the network, txt or bin? after unpacking do they look the same in memory?
and I wonder why, after connecting networkg g170-b15c192-s74759936-d68801237.txt.gz, KataGo is initialized for 22 seconds, and after network g170e-b15c192-s1672170752-d466197061.txt.gz for 4 seconds? tried many times

go4thewin wrote:Amazing speed increase. With the new optimized net on my snapdragon 625, with the default settings it plays at 9d bot strength, even if you reduce it to just 1 thread!

network remains the same, its use is optimized

And wrote:does it matter in what format the network, txt or bin? after unpacking do they look the same in memory?
and I wonder why, after connecting networkg g170-b15c192-s74759936-d68801237.txt.gz, KataGo is initialized for 22 seconds, and after network g170e-b15c192-s1672170752-d466197061.txt.gz for 4 seconds? tried many times

In my understanding they will look the same in memory. But .bin.gz loads faster, so if you have both formats available, you should prefer .bin.gz

The 15b models you quoted have the same loading time on my handset, so I have no explanation for the difference on your device.

@And, please please please get yourself this for christmas and test out if the npu feature in baduk ai works https://www.amazon.com/dp/B0824QJHKW?ref=myi_title_dp

that would be crazy if we start getting huge numbers of playouts on $100 devices. i already am getting more than i could ever use!

even on my $220 rockchip rk3365 e ink device the 20b net can make instant moves now.

akigo wrote:I made a new version (0.18) with the following change. Furthermore it will try to use the NPU if possible. The effect depends very much on your device. On my handset with Snapdragon 855+ it is about 6 times faster than the previous version.

The second change is quite experimental as it is very HW-dependent and I could test it only on my device. If you experience problems, please tell me.

thanks so much akigo! so if i understand correctly, if i plug an npu stick https://www.amazon.com/Orange-Pi-Neural ... ACXX9CH5E2 in the otg port of my ereader, it may increase playouts? wow

also does windows/linux katago builds support npu?

go4thewin wrote:@And, please please please get yourself this for christmas and test out if the npu feature in baduk ai works https://www.amazon.com/dp/B0824QJHKW?ref=myi_title_dp

that would be crazy if we start getting huge numbers of playouts on $100 devices. i already am getting more than i could ever use!

even on my $220 rockchip rk3365 e ink device the 20b net can make instant moves now.

Thanks, but ... I can't find time to check how KataGo works on my netbook with an AMD C-60, and install the GF GTX 1650 video card (bought back in the spring!) in a computer with Ryzen 5 and see how KataGo works there!

thanks so much akigo! so if i understand correctly, if i plug an npu stick https://www.amazon.com/Orange-Pi-Neural ... ACXX9CH5E2 in the otg port of my ereader, it may increase playouts? wow

Unfortunately not, as the stick can not be accessed by standard android interfaces. It has its own SDK, that is you must write programs specifically for the stick.

also does windows/linux katago builds support npu?

No, the NPU access I added is android-specific. On a PC the way to accelerate is still a powerful graphics card.

oh well i guess its already faster than i could ever need any chance to add gnugo level 1-10?

go4thewin wrote:oh well i guess its already faster than i could ever need any chance to add gnugo level 1-10?

net 6b(b6c96-s34290432-d5157105) 1p - gnugo 3.6(12k-13k) 1:4
net 6b(b6c96-s24455424-d3879081) 1p - gnugo 3.0(16k) 3:3
from networks 6b can you probably choose the one suitable for the game?

@And, thanks but they both sometimes play 1st move on 2nd line. I think for that level 6b s114 with kyu_rank 7 is better? s114 with 1po is like a human a little bit.

why not choose a position to your liking and continue playing with KataGo?

this seems to work well. if you pre place the stones on the star points like the older style of playing go http://gokifu.com/s/2i9-gokifu-17000101 ... shi().html then it seems like a fun game with s342 if you don't want to use the calibrated ai. I wish you could just get it to use an opening book for the first moves.
EDIT: just found out you can play the opening moves with the bigger net, then use the smaller net and hit the "load last game" button in Baduk Ai and continue your game. Thanks And, great idea