search

Yikai-Wang / ICI-FSL

This repository contains the code for Instance Credibility Inference (ICI).
82 stars 22 forks source link

Better experimental results on the CIFAR-FS dataset #14

Closed lampardwk closed 3 years ago

lampardwk commented 3 years ago

Many thanks for making the latest code available. I am trying to run the experiments from scratch, but I am getting around +4% more accuracy on the CIFAR-FS dataset while I get similar accuracy for the rest of the datasets. The parameters in all experiments are strictly based on the description in the paper. I was wondering if this is normal and what may cause this problem. Thanks in advance.

Yikai-Wang commented 3 years ago

Hi there, thanks for the attention to our work. As you can reproduce experiments on the other datasets, the only reason seems to be the dataset. Did you try to use different random seed? Compared to other datasets, CIFAR-FS is a low resolution image dataset, which may be more sensitive to random seeds. Besides, please note that we resize all the images into 84 X 84 pixels. Hope it can help you.

lampardwk commented 3 years ago

Many thanks on your quick reply. I'll repeat the experiment according to your suggestion.