New android app "BadukAI"

[akigo]

This can't be caused by the type of network since the file dialog doesn't look into the files. It just shows all filenames which end with .bin.gz or .txt.gz

So either:
1) The transfer from PC somehow changed the file extension
or
2) BadukAI doesn't have permission to read the folder where you put the network

You could try downloading the network directly from the "Network File" column on katagotraining.org/networks via browser on your phone. Then go to "Download" folder in BadukAI's model selection dialog. It should be there ...

Furthermore BadukAI may have no "Storage access" permission (after uninstalling and installing a new version, you have to grant this again). How to check that depends a little on the android version, but usually a long press on the app icon brings up a context dialog with an item like "details". After tapping that, a lot of information about the app is displayed, including permissions.

[Zeed]

Yeah, I tried to download the file from the site and it worked. And I also figured that if i rename it to smth like "Kg40b.gz" it doesn't work but it doesn when it's "Kg40b.bin.gz", so that ".bin" thing seems to be essential.
Anyways, now it all works perfectly fine, thank u veru much!

[Zeed]

I actually just realized that u already told me that there needs to be .bin, I somehow missed that. Thanks again for such a detail answer.
And one more question, I hope I am not repeating it. Those nets I download myself are not getting optimized, aren't they? 20b works way faster than 15b and only a bit slower than 6b. I just want to prove whenever it's normal or I can make them work faster somehow.

[akigo]

And one more question, I hope I am not repeating it. Those nets I download myself are not getting optimized, aren't they? 20b works way faster than 15b and only a bit slower than 6b. I just want to prove whenever it's normal or I can make them work faster somehow.

You are right, nets you download yourself are not optimized. That's because the optimization takes a long time, is not completely automatic and needs tools that are not running on android. So unfortunately it can't be done "on the fly" in the app.

[lovego]

Can you provide scripts to optimize weight? So we can change the weight we like.

[lovego]

And one more question, I hope I am not repeating it. Those nets I download myself are not getting optimized, aren't they? 20b works way faster than 15b and only a bit slower than 6b. I just want to prove whenever it's normal or I can make them work faster somehow.

You are right, nets you download yourself are not optimized. That's because the optimization takes a long time, is not completely automatic and needs tools that are not running on android. So unfortunately it can't be done "on the fly" in the app.

Can you provide scripts to optimize weight? So we can change the weight we like.

[akigo]

Can you provide scripts to optimize weight? So we can change the weight we like.

It's not that simple, because:

1) The optimization needs a complex environment of third-party tools. I built that for myself over a couple of months with a lot of "trial and error" diversions. So I couldn't even give a straightforward recipe how to get there from scratch (to do that, I would have to restart from scratch again myself, and I don't have the time to do that).

2) These tools have some bugs, that have to be corrected manually during the optimization process, so anybody doing this can't just run a script, he has to understand exactly what's going on.

3) The optimization is different for each type of network, e.g. for the new KataGo 40b ones it doesn't exist yet, I still have to develop it.

[lightvector]

I'm curious - is this some sort of quantization plus fine-tuning using data to limit the accuracy loss due to the quantization?

[akigo]

I'm curious - is this some sort of quantization plus fine-tuning using data to limit the accuracy loss due to the quantization?

Yes, that perfectly sums it up in one sentence.

[lovego]

Can you provide the optimized weight of the latest katago 40b s580
I think it must be better than the current 20b

[akigo]

Can you provide the optimized weight of the latest katago 40b s580
I think it must be better than the current 20b

Not yet (see part (3) of my answer to your last post), but it's on my todo list.

[akigo]

I made a new preliminary version (1.4), up to now only available on github

https://github.com/aki65/aki65.github.i ... tag/v1.4.0
or
https://aki65.github.io

It is a little experimental as I added a new network acceleration method which kicks in if the existing one doesn't work to full extent because the required hardware is not available. Together with 32/64 bit this makes up for a lot of possibilities (depending on device type) which I can't check all on my own device, so I kindly ask for your help:

@go4thewin
From your earlier posts I know that you have a 64 bit device where the existing acceleration fully worked. Could you please check if it still works as before with the new version (KG and LZ) ?

@And
You own a 32 bit device where the existing acceleration fully worked. Could you also please check if the new version works as before ?

@Heinz-Dieter Conradi
From your github posts I know that yown a OnePlus 8 Pro where the existing acceleration did NOT work to full extent (only 50% better than original). On your device type the new method should improve speed significantly (for both LZ and KG). Could you please check if this really works ?

If all tests go well, I will make the new version official and put it on play store.

[go4thewin]

perfect as always for me, instant moves with 1 playout snapdragon 625. i think it may have gotten even faster, thanks!

[lovego]

Great app! thank you very much

[And]

on tablet (MT8735M) KG 20b works also fast (~ 50 visits/min from starting position)
here https://www.mediatek.com/products/tablets/mt8735m it is written that the processor is 64bit, but the 64bit version is not installed. probably android 32 bit
I'll check for mt8321 today

and thanks again !!!