train-gammas.sh strange number of games

[dsmic]

I remember having used this successfully. By the way, there is a line where I changed the probability to 0.1 to take a move.

Now I do not see the correct number of games from the mm tools and they are much to quick?

If I load 1315 games I see Games=230 from mm-tools?

Do you have any idea what happend?

I saw this by training 13x13 games to get a feeling of the behavior of the gammas. I attach the games just for reference.

[Francois van Niekerk]

This does seem weird. The number of games reported by the mm tool is the number of bradley-terry competitions, so it won't be the same as the number of games, but that number seems very low.

Are you running train-gammas.sh directly? Are you sending it the entire list of 1315 games? You can try revert train-gammas.sh to a prior state to make sure a recent change hasn't broken something

[dsmic]

reverted to version of 2012-09-16 and behaves quite different. It now takes some time after extracting competions is printed:)

I use command ls ../../../games//.sgf|./train-gammas.sh initial.gamma large > trained.gamma

where in games/Name_of_pro/*.sgf the pro games are

[Francois van Niekerk]

Do you mean that the version of 2012-09-16 reports a correct number of competitions?

Which version are you having the problems with?

[dsmic]

2012-09-16 is still thinking, that is good for 1315 games, as far as I remember.

I tried you latest version and my latest version, both did not work correctly. Your latest even showed a lower number of competitions. Both version run within seconds to the end...

I will keep you informed, as soon as I get a result from the run. It will take a while, possibly it will even crash due to memory consumption, as I run it with probability 1.0. If I remember correctly I had to reduce the probability to get in im my 8GB.

[dsmic]

The version of 2012-09-16 worked. I trained with 13x13 games oakfoam against gnugo level 10.

Interestingly the playing strength of the new parameters is slightly weaker (but close to statistical significance 3000games 54% against 56.3% winrate)

[dsmic]

sorry, I did not search for the bug up to now, I hate sh scripts:)

[Francois van Niekerk]

Do you only get this problem with 13x13 games? Do 19x19 games work fine?

[dsmic]

No, did not work for 19x19 too.

I only did not finish the test run with 19x19 and 1315 games and the 2012-09-16 version as I did not have enough hard disk space, ram and time at the moment.

But it seemed to work. The newer for sure did not work, they finisched after seconds in the mm-tools with much to low number of games and very bad gammas.

[Francois van Niekerk]

Could you check the how many competitions are gathered by the script? It should be as simply as adding:

cat $TEMPMM | grep '#' | wc -l

before the MM tool is called.

[dsmic]

this is the file, with your latest version of train-gamma.sh it has 26 # lines

1315 games were loaded. With 2012-09-16 its working, but I can not tell you the number very quickly, takes a long time...

[dsmic]

After merging in, I still have this problem. As you probably do not have it, there must be some script incompatibility or I do somethin wrong.

Can you give me a sample call with sample games which works for you (e.g. two .sgf and your initial gamma) to see what I get?

[dsmic]

The version of gogui should be 1.4.7 or later, I had 1.4.3 (this needs undo availible in the engine)

Sorry for the noise

[Francois van Niekerk]

No problem. Glad it is resolved :)