In this post I would like to explain math used in eval_count_of_generated_ft() function proposed as a part of new interactions generation mechanism to VW 7.10.1 master branch.
Notes on Machine Learning
суббота, 16 мая 2015 г.
воскресенье, 16 ноября 2014 г.
Fixing 'best constant' and 'best constant loss' in VW
It's well-known that best constant value that printed by VW in work summary is correct only for squared loss. I have also pointed out that best constant loss value is never printed out for classes other than {0,1} and John Langford suggested anyone to try to fix this or he may remove it at all otherwise.
So, I have decided to try. Below is quite scrupulous description of what and why have I done in code so anyone could easily verify my conclusions. I nigher ML nor mapple expert so such verification really has sense.
So, I have decided to try. Below is quite scrupulous description of what and why have I done in code so anyone could easily verify my conclusions. I nigher ML nor mapple expert so such verification really has sense.
суббота, 15 ноября 2014 г.
Debugging VW with QtCreator on Linux
While playing with Vowpal Wabbit I have struggled many times with impossibility to debug VW sources in IDE and check how exactly it works. As I've recently finally switched to Kubuntu I don't want to launch Win every time I'm facing small problems. And unfortunately VWs sources equipped with Visual Studio project files which doesn't support *nix platforms.
Finally I've forced myself to try building VW on Linux as Qt project with QtCeator IDE. Surprisingly it was very easy to do in just a few steps:
Finally I've forced myself to try building VW on Linux as Qt project with QtCeator IDE. Surprisingly it was very easy to do in just a few steps:
Подписаться на:
Сообщения (Atom)