Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
  • Jul 15 17:35
    rey-allan starred ujh/iomrascalai
  • Dec 12 2019 03:54
    hbzhouwei starred ujh/iomrascalai
  • Jul 13 2019 04:57
    adityasadhu starred ujh/iomrascalai
  • Mar 25 2019 00:00
    fgonza52 starred ujh/iomrascalai
  • Jan 23 2019 23:01
    rudi3838 starred ujh/iomrascalai
  • Sep 24 2018 04:58
    ankworld starred ujh/iomrascalai
  • Jun 12 2018 01:01
    Terkwood starred ujh/iomrascalai
  • Jun 04 2018 11:16
    esneko starred ujh/iomrascalai
  • Mar 07 2018 20:25
    ufobat starred ujh/iomrascalai
  • Jan 06 2018 13:36

    ujh on refactorings

    (compare)

  • Jan 06 2018 13:36

    ujh on master

    Remove code to keep track of de… Don't simulate more than 10 thr… Merge pull request #327 from uj… and 8 more (compare)

  • Jan 06 2018 13:36
    ujh closed #328
  • Jan 06 2018 13:36
    ujh synchronize #328
  • Jan 06 2018 13:36

    ujh on refactorings

    Note performance in changelog (compare)

  • Jan 06 2018 13:34
    ujh commented #328
  • Jan 06 2018 13:33
    ujh commented #328
  • Dec 29 2017 06:28
    clearlycloudy starred ujh/iomrascalai
  • Dec 21 2017 11:45
    ujh edited #328
  • Dec 21 2017 11:45
    ujh commented #328
  • Dec 21 2017 11:43
    ujh unlabeled #328
iopq
@iopq
I kind of forgot where to find this, what's the easiest way to print out the amount of playouts each coordinate has?
Urban Hafner
@ujh
I'm not entirely sure what you mean my that statement. Are you referring to how often each child of the root was picked for a playout? Then you just need to print out the plays field of the Node struct.
iopq
@iopq
so I just need to grab the root somewhere and go over every child?
Urban Hafner
@ujh
Exactly. All coordinates should be there.
Urban Hafner
@ujh
So, we didn’t come last in yesterday’s tournament: http://www.gokgs.com/tournEntrants.jsp?sort=s&id=1030. And more importantly we now know what is still missing so that we can participate without crashes in the next tournament. Apparently implementing the kgs-chat command causes the crash. I had implemented the command as I assumed it was needed to avoid the crashes. Oh, well. It’s easy enough to fix. :)
iopq
@iopq
woot
Urban Hafner
@ujh
I also may have some time tomorrow to work on the bot. Have you started implementing reg_genmove or should I have a go?
iopq
@iopq
haven't done it yet, I'm doing uct_gfx
Urban Hafner
@ujh
OK, so I’ll have a got at it then.
iopq
@iopq
what is this wizardry? running iomrascalai by itself gives me the error that it can't find libcaffe.so, but cargo run --release runs fine
Urban Hafner
@ujh
That is odd. It works for me. But it may be necessary to define some library search paths.
iopq
@iopq
I think cargo is smart about those, I had to put the .so file in my lib64 folder for it to work out of gogui
but I did genmove and it's taking infinity time to finish
Urban Hafner
@ujh
Yes, it's horribly slow.
And I'm not even sure that it's correct. 😜
iopq
@iopq
aren't you only supposed to do it on the root nodes
it's been 5 minutes since I asked for a genmove on a 19x19
Urban Hafner
@ujh
Yes, I will try that next. But I also want to figure out if I'm even getting the right data from the NN.
But it's all just a prototype to see if something like this can work.
iopq
@iopq
well since I can't get it to give me ANY move I can't tell if it works
Urban Hafner
@ujh
That is really odd though.
iopq
@iopq
it's been 12 minutes since I asked for a genmove
Urban Hafner
@ujh
On my machine it actually respects the time limits. Really odd.
iopq
@iopq
it's running in a VM
iopq
@iopq
it's been 40 minutes, I'm starting to think I'm never getting a move
I recompiled it to use the files from the same folder
iopq
@iopq
I'm still not getting a move back even with the latest version
Urban Hafner
@ujh
That’s weird. Sorry about that. But this will never be merged into master anyway. The NN I used is only useful for 19x19, caffe is really slow and the Rust caffe bindings are a giant hack anyway. What I’m trying here is just to get a sense of the tradeoffs involved.
I just read the Facebook darkforest paper and it seems that implementing something like that from scratch seems more useful. At least I will know how the NN actually behaves then. But I guess that’s a job that will take a few months. So maybe implementing easier stuff (like the large patterns from michi) first would be better.
iopq
@iopq
that's strange because NN with no search should easily beat gnugo
on a 19x19 board
Urban Hafner
@ujh
Yes, it's all very weird. Maybe it's the particular NN? It only has a 44% prediction rate. Or maybe I'm just using it wrong?
Oh well, I wanted to learn how to train my own net anyway. 😁
iopq
@iopq
I mean, does it play like this? https://chrisc36.github.io/deep-go/
Urban Hafner
@ujh
I was going to read that paper next. :)
Urban Hafner
@ujh
But I'll have a look at the Michi large patterns first. Maybe I can get that done before the next tournament.