Life In 19x19

Limeztone wrote: your app is set to use KataGo and I open a game stored in sgf, the Win Rate totally off. It seams that it somehow is affected by the end position of that game. And it keeps being wrong if you step through the game (I am surprised that no one else has stumbled on this bug). The bug is not in just this version. It was in the previous versions as well.

Is it really only for me that this happens?
If someone care to verify you can use the Sample.sgf that I included in my previous post.

y_ich san, nice redesign of undo! And thanks for confirmation of pass!

Limeztone wrote:

Limeztone wrote: your app is set to use KataGo and I open a game stored in sgf, the Win Rate totally off. It seams that it somehow is affected by the end position of that game. And it keeps being wrong if you step through the game (I am surprised that no one else has stumbled on this bug). The bug is not in just this version. It was in the previous versions as well.

Is it really only for me that this happens?
If someone care to verify you can use the Sample.sgf that I included in my previous post.

y_ich san, nice redesign of undo! And thanks for confirmation of pass!

This issue exist in the newer versions as well.
It's a bit sad not to be able to "review" sgf-games...

Hi y_ich,

Still love your work, and hope you are doing ok in the virus crisis.

Any chance that you can update the app with the final katago models now that it’s completed it’s most recent run?

Thanks!

y_ish san,

please let us know if you still read this forum!

I have a tiny suggestion!
At least I think it would be very easy to implement.
It's quite normal to play with reversed komi.
And as I see there is today no way to set up you app with reverse komi (other than -0.5)
So we can't study with reversed komi, if we want.

Thanks in advance!

Limeztone wrote:y_ish san,

please let us know if you still read this forum!

...

You can always hit the "profile" button below any of his (or anyone else's) post and check when they were last on L19. As I write this y_ish was last here on Friday, May 29th (Japan Standard Time).

For anyone who didn’t know, y_ich did update “A Master of Go” with the final katago models.

On my base iPad and iPhone 11Pro (both with neural engine) the 40 block engine gets around 20 playouts per second, and the 20 block gets 40 playouts per second. This is good enough that my phone can find some solutions that Leela Zero can’t find on my computer. Amazing.

Y_ich, thanks, and please come back.

wineandgolover wrote:Hi y_ich,

Still love your work, and hope you are doing ok in the virus crisis.

Any chance that you can update the app with the final katago models now that it’s completed it’s most recent run?

Thanks!

Just announced: OpenML speed increased 10x on new iPad Air.

If that transfers to A Master of Go as past gains have, katago's current 40b engine might get around 200 playouts per second.

wineandgolover wrote:Just announced: OpenML speed increased 10x on new iPad Air.

If that transfers to A Master of Go as past gains have, katago's current 40b engine might get around 200 playouts per second.

If I have understood goame correctly, 2x RTX 2080 TI do 1358 playouts per second at 2x 258W so 2.5 playouts/Ws. Apple A14 with 200 playouts per second and ca. 6.2W would do 32 playouts/Ws, i.e., ca. 13 times as power-efficient but 6.78 times slower.

Hello, everyone.

Since my last post I haven't received any notification e-mails, I haven't checked this thread.
I will read all posts after that and I will check notification settings here.

I must let you know that KataGo engine on "A Master of Go" does not work on iOS 14/iPadOS 14.
Please stay in iOS 13/iPadOS 13 until I fix the issue.

Thank you for your interest on my app!

Limeztone san,

I am sorry for my too late reply.
Still I could not reproduce your issue by Sample.sgf.
Sample.sgf sets komi to -5.5 and KataGo suggests 11 points lead for black.
Could you tell me the meaning of "the Win Rate totally off"?

Thank you for your cooperations!

wineandgolover san, RobertJasiek san,

"10x speed up" means that A14 has an AMX unit, which already introduced on A13.
This unit seems not effective for inference. It is for learning.

A14 has Neural Engine which runs twice faster than A13.
I expect that A14's eps(evaluation per second) will be nearly double of A13 for Leela Zero-type weights.
But I don't expect such a boost for KataGo weight by A14 since its bottleneck is not convolutions, is global pooling layers.


Please take care of yourselves in COVID-19 era.

y_ich wrote:Hello, everyone.

Since my last post I haven't received any notification e-mails, I haven't checked this thread.
I will read all posts after that and I will check notification settings here.

I must let you know that KataGo engine on "A Master of Go" does not work on iOS 14/iPadOS 14.
Please stay in iOS 13/iPadOS 13 until I fix the issue.

Thank you for your interest on my app!

Damn, I just came here to say this. Guess I'll wait for the fix. Good luck, and thanks y_ich!

wineandgolover san,

I am sorry that you have already updated iPadOS.

In my investigation the cause is inside CoreML(part of OS).
I reported it to Apple.
I also asked any workarounds but don't expect it so much.
I hope that Apple will update OS with fixing soon.

Grünauer san,

Thank you for your report!

Could you send it via "Bug report/Feature request" link in the main menu?
It adds related information to help me know issues in detail.
This is the official, highly recommended way.

I could not reproduce your issue in version 5.6.0, which was released on 17:00 Sep. 20th UTC.

Thank you for your cooperations.

Dear everyone,

As Grünauer san kindly reported it, fixing the issue of KataGo on iOS 14/iPadOS 14 by the version 5.6.0 was not complete.
If your device does not have Neural Engine, please stay in iOS 13/iPadOS 13.
If your device has Neural Engine, I think that it works hopefully.

Thank you for your patience.