notani / python-glad Goto Github PK

Hi, I was excited to find your Whitehill GLAD implementation and played a little bit with it. I found that multiclassification variant glad.py needs some fixes to be able to run. I made it but generally there a couple of problems:

• it works seriously slow, like 10 hours on server for 200k source labels. It somehow crashes on full dataset with 1.3M labels
• some warnings for exponent overflow and division by zero

Generally after the calculation I am not sure that I got some correct results. Labels have a lots classes with equal probabilities. Alpha distibution has lots of negative values while default is 1.0, beta a lot of with >>2.0. Not sure if it is right.

I can see that a lot of efforts have been put into this implementation so I guess there is a more correct version somewhere that is not commited to github? Could you please comment on this?

Thank you a lot!