minigo icon indicating copy to clipboard operation
minigo copied to clipboard
verified publisher icon tensorflow

Test for "symmetry divergence" during training

Open brilee opened this issue 7 years ago • 2 comments

One of the symptoms of value net overfitting that we found was that the 8 board symmetries would yield wildly different results when put through the value net. So instead of the value net returning roughly consistent estimations, stdev=0.1 or so, it would instead disagree completely, with 7 symmetries saying that B+ 0.99, and the 8th symmetry saying W+ 0.99.

It'd be nice to log the average stdev of the 8 symmetries (or some other characterization of how divergent the symmetries are.)

brilee avatar Feb 12 '18 20:02 brilee

So a script to analyze these is in #139. @sethtroisi do you think this makes sense as a metric to watch in TB, or should we put this in cloudygo?

amj avatar May 31 '18 18:05 amj

It's harder to add to CloudyGo than to TB.

I'm not sure which is more semantically correct

sethtroisi avatar May 31 '18 19:05 sethtroisi