Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
  • May 05 2020 11:04
    wrongu closed #233
  • May 05 2020 02:24
    ifhbdod opened #233
  • Feb 27 2020 22:10
    renata2216 opened #232
  • Feb 27 2020 20:01
    gromovadarya90 opened #231
  • Dec 02 2019 13:39

    wrongu on master

    removed large model_test.zip fi… removing further intermediate h… (compare)

  • Dec 01 2019 23:58

    wrongu on master

    Adding a partially trained CNNP… Weights from 100n epochs in 'fu… training session KGS 6D+ data … and 17 more (compare)

  • Dec 01 2019 22:33

    wrongu on master

    removed large model_test.zip fi… (compare)

  • Dec 01 2019 22:32

    wrongu on master

    removing many intermediate trai… removing expert play files to f… Merge remote-tracking branch 'o… and 1 more (compare)

  • Jun 11 2018 12:50
    wrongu closed #229
  • Jun 11 2018 12:50
    wrongu commented #229
  • Jun 11 2018 05:25
    CezCz commented #229
  • Jun 09 2018 13:55

    wrongu on develop

    Fixed Travis-CI Theano build, u… (compare)

  • Jun 09 2018 13:37
    wrongu commented #229
  • Jun 09 2018 07:58
    CezCz commented #229
  • Jun 08 2018 17:24
    wrongu commented #229
  • Jun 08 2018 17:23
    wrongu closed #230
  • Jun 08 2018 17:23
    wrongu commented #230
  • Jun 08 2018 16:44
    782832949 opened #230
  • Jun 08 2018 16:09

    wrongu on develop

    cython-v1-BETA added boardsize warning Merge pull request #209 from Ma… and 43 more (compare)

  • Jun 08 2018 16:03
    CezCz commented #229
jackieli19
@jackieli19
Oh, i think i know what that means in english. "B+R" (Black wins by resign) or "B+3.5" (black wins by 3.5 moku).
oriskunk
@oriskunk
@MaMiFreak Hi.. : ) Are you planning to implement 17 features for zero use?
jackieli19
@jackieli19
@oriskunk thanks for your information!
oriskunk
@oriskunk
:smile:
MaMiFreak
@MaMiFreak

@jackieli19 indeed that is the correct interpretation

@oriskunk i found a way to generate them by adding a parent variable to gamestate but it is not completely compatible with our current code. right now i have an implementation that always creates a parent by copying the gamestate before playing a move but in some cases it slows down the code without a reason ( slows down ), still have to think of an elegant solution for this.
i ran a test and i seems that the agz input works just as good as board + turns_since so i started training my agz 9*9 network with those features. it still can't beat a weak leela but it is definitely improving.

jackieli19
@jackieli19
@MaMiFreak I am playing around with your combined scripts. I find that the time per Go Game has a sudden drop at the beginning several epochs as the version increases whether for the white side or the black side. It is really interesting. As for the network policy, its time per game decreases make sense. But the white side is the leela or Pachi, it should be stable in my understanding. I am curious about this phenomenon.
image.png
jackieli19
@jackieli19
another thing i am confused is that we reach a similar accuracy as AlphaGo (57%) using a much fewer sgf files compared to 30 million positions that AlphaGo trained. The training takes approximately 2 months. Is it a really goof replication of AlphaGo's policy training phase? Or the high accuracy is something resulting from a strong overfitting... ememem
MaMiFreak
@MaMiFreak

@jackieli19 why does it make sense that pachi should use the same time and the network gets faster?
i think you have to calculate thinking time relative to amount of moves in a game because games played by first generation are on average 400+ moves per game and at the last generation around ~190 moves per game. i think that is exactly what your graph shows

good question, first about your assumption on the amount of sgf used, why do you think i used less sgf? ( i have to check exactly what version i used but i have a 27M, 34M and a 40M positions set. i think i used the 27M set because that is also the set i used to train the previous model )
Is it a good replication of the training alphago, i think yes but i had to make some changes in order to keep the training time manageable. they trained every learning rate for 80M batches, i switched to the next learning rate when it was no longer improving and thus not training nearly the 340M batches they trained. they averaged weights over the last 100 updates. at some point i started training with batches of 160 to somewhat get the same effect but i think the accuracy would be higher if i started with a higher batch size from the beginning. those differences resulting in a 1% reduction in accuracy compared to AlphaGo ( they reached 58% ) and yes that last percent would increase playing strength with almost one rank( according to the AG paper )
as for overfitting, it would mean the model learns what to do in certain situations and not able to generalize to new situations, you can check all the games it played with the dataset but i am pretty sure there are a lot of positions that are not in the dataset and still the net is able to play good moves and win the game.
i only did a small test with pro game positions not in the kgs dataset to confirm this.

let me know if you disagree or have more questions=)
MaMiFreak
@MaMiFreak
one more thing on the game time, i run test on several machines and i am not completely sure all of them have been done on the same machine. i am pretty sure several of the leela games have been done on a different one but i am not sure about the pachi games..
jackieli19
@jackieli19
@MaMiFreak Thanks for your reply.
for the game time, I understand the decreasing time is due to the less moves per game. But I am confused why the average lengths of a game decreases from 400 to 190.
For the data set, is it because of my misunderstanding of the definition of positions. I think one move is one position. So we have two data sets in our data repository. One is the "Community" another one is "kgs". I think you use the latter for your 192 filters network. We have 554 sgf files in total with roughly 300 moves per game, which add up to 166k positions ?
jackieli19
@jackieli19
Oh, I think maybe you use the Communty folder as well....
MaMiFreak
@MaMiFreak

@jackieli19 have a look at the sgf, the later generation games play a decent game, not to many stones being captured etc. often resuting in pachi resigning. the first generations is a bunch of moves played by the net huge groups dying and our net keeps on playing in those places until there are no more viable moves

ah yes i understand, i have a kgs dataset and managed to get two updates but they asked me not to share the sgf publicly so i generated a hdf5 file of it and shared it here, i mentioned it here but that was quite a while ago

jackieli19
@jackieli19
Ohhhh, thanks....I got it..I should notice this sentence "We have a few datasets of professional play in SGF format, not all of which we have permission to share."
@MaMiFreak thanks for your explanation ^^
MaMiFreak
@MaMiFreak
you're welcome=)
MaMiFreak
@MaMiFreak
nice update on 9x9 AGz training, generation 44 managed to win the first games versus leela! it took only 16500 games..
oriskunk
@oriskunk
@MaMiFreak Can you tell me more about AGz?
  • Does it use the same method as self-play of Zero?
  • Number of playouts to use for one move
  • Number of games required for one update
  • The average time taken for one move. Average game time
:smile:
jackieli19
@jackieli19
@MaMiFreak i guess you trained the network in the personal computer. I am wondering whether we can make use of the google cloud computing. Will it improve the training time?
jackieli19
@jackieli19
the 192 filter network approaches a 50% winning ratio vs Leela 0100, can we say our strongest policy has a similar level as leela which is a 4-dan amateur level?
MaMiFreak
@MaMiFreak

@oriskunk of course:

  • yes but i start the strength comparison after training a new generation manually because i train again after a small amount of games and i have to check for signs of overfitting
  • 400 playouts per move (during training both players use the same gametree)
  • up so far i trained every 500 games, but right now it looks like i have to increase that number to avoid overfitting, i am in doubt if i should try to play 50% of the games with 100playouts and 50% of the games with 400 playouts in order to get more data quicker. i don't know what is more important, quality or quantity..
  • on average ( with strength testing ) i generate a 1000 games per day, due to bug hunting and the parallel implementation i have not yet checked time per game/move because i'd also have to figure out how long each game has to wait in queue for nn prediction

@jackieli19

  • yes and yes but it will cost a dime and code has to be optimized for it
  • no because leela has a playout setting and the one i used is a very limited one, any leela ranks you can find have to be with the same settings to be comparable. @robert-waite is running the newest net on kgs and it looks like it might be around 1d/2d but that is still an estimate
jackieli19
@jackieli19
@MaMiFreak thanks~
oriskunk
@oriskunk
@MaMiFreak thx. :smile:
MaMiFreak
@MaMiFreak
MaMiFreak
@MaMiFreak
for who wants to download the sgf used: https://alphagoteach.deepmind.com/dist/files/book.sgf
MaMiFreak
@MaMiFreak
@robert-waite i was looking at the NeuralZ01-NeuralZ05 games, are they all playing 57% policy net without any look ahead or are there differences?
robert-waite
@robert-waite
All were the 57 net.. just single greedy eval. No diff between bots
oriskunk
@oriskunk
Do you know CGOS?
http://www.yss-aya.com/cgos/
It seems to be good for evaluating bots.
You can see the ranking with Elo, and it's easy to work with.
This is not an advertisement that our bot is banned from KGS.
To T
vftens
@vftens
@MaMiFreak I've solved last error in gtp_wrapper in Windows 10 in Python 3.6.3. So tests works even on weak videocard 810M, even without Anaconda 3 in Windows 10. Results are in https://github.com/vftens/RocAlphaGo/tree/aug25-keras2-py35 . I'm using there latest versions of Keras and TensorFlow-gpu==1.4.0 2) All other testing continues, as far as I understand the subject.
wrongu
@wrongu
Hey all, this is long overdue, but I spent this week rewriting @MaMiFreak’s cython code.
This started a while ago as just small "stylistic" cleanup of the cython branch to get it ready to merge into the main project, but eventually grew pretty big. The goal is to have fast core game logic with consistent style, readability, and maintainability. Highlights include:
  • refactored the massive AlphaGo/go.pyx and AlphaGo/go_data.pyx files into an AlphaGo/go/ package.
  • inside this package the core game logic is split into multiple smaller files
  • all C++ memory management with smart pointers (no more malloc and free)
    There's still some work left to do here, but at this stage I'd like to get feedback from anyone who is willing to take a look. Code is here
Patrick Wieschollek
@PatWie
This might be relevant as well: https://github.com/cgtuebingen/tensorpack-recipes/tree/master/AlphaGo which is written entirely in c++ and provides python-bindings. Feel free to copy parts, if needed or helpful
wrongu
@wrongu
Thanks!
Codist
@countstarlight
I'm trying to transplant RocAlphaGo to play Game of Amazons, and there are problems when trying to implement supervised policy train.
The error is: ValueError: Error when checking model target: expected activation_1 to have shape (None, 60) but got array with shape (10, 100)
The train data set is a (10, 6, 10, 10) array. How to solve it?
Gaoyang Tang
@tygrer
I use python3.6. The go_python.py file includes long type. So how can I change to the int type that the program runs correctly?
vftens
@vftens
@tygrer What's wrong with this version https://github.com/vftens/RocAlphaGo/tree/aug25-keras2-py35 ?
Gaoyang Tang
@tygrer
It can run. Thanks @vftens !
vftens
@vftens
Tested the stuff with latest Tensorflow==1.8.0 with keras==2.2.0 with CUDA 9.0 and cuDNN 7.1.4 on Windows 10 Python 3.6.5
Abhijeet Go-kar
@abunickabhi
Nice!
oriskunk
@oriskunk
is anyone attending ewrl conference?
Abhijeet Go-kar
@abunickabhi
Nope not attending that conference
xiapo00
@xiapo00
'Tis wonderful!