search

abhijitbendale / OSDN

Code and data for the research paper "Towards Open Set Deep Networks" A Bendale, T Boult, CVPR 2016
Other
245 stars 57 forks source link

About choosing the hyper-parameters #5

Closed yanxiangyi closed 6 years ago

yanxiangyi commented 6 years ago

Hi! @abhijitbendale Could you please share your experience about choosing the hyper parameters, such as ALPHA_RANK and WEIBULL_TAIL_SIZE?

yanxiangyi commented 6 years ago

Emailed the author. He replied as follows.

Hi Xiangyi Yan,

Thanks for your interest. For ALPHA_RANK, I tried few iterations and rank 10 seemed to be performing best across categories. For WEIBULL_TAIL_SIZE it is a bit tricky. More specifically it depends on the total size of data you have for fitting. For imagenet, each category has about 700-1000 images. Hence if you consider tail or a quantile of distritbuion it is about 70-100 images. However, weibull distributions are more effective in predicting extreme events, thats what you need even smaller tail-sizes to harness power of the distribution. Thats why I considered even smaller tail sizes. Very small tailsize (less than 10) dont give enough data for fitting. Too large tail sizes (more than 40 with data distribution of 1000) dont give enough variability.

Abhijit