search

accosmin / nano

C++ library [machine learning & numerical optimization] - superseeded by libnano
MIT License
1 stars 0 forks source link

Feasible benchmark of the trainers #57

Closed accosmin closed 8 years ago

accosmin commented 9 years ago

Should increase the number of samples to more consistent results. Should investigate a fairer and faster minibatch setting.

Ideally this benchmark should not take more than 2-3 hours on a decent 4+ core system.

accosmin commented 9 years ago

Currently a minibatch run takes around 25-40 times more than a batch one. This difference should be greatly reduced. Also the regularized cases will make the processing even slower (e.g. ~10x).