search

yangyangyang127 / APE

[ICCV 2023] Code for "Not All Features Matter: Enhancing Few-shot CLIP with Adaptive Prior Refinement"
135 stars 12 forks source link

Few questions #1

Open vishaal27 opened 1 year ago

vishaal27 commented 1 year ago

Very interesting work and great results in both training-free and training-required regimes. I had two questions:

  1. I was wondering about a random selection baseline -- how effective would that be by just randomly subsampling the feature selection indices without using APE. I see there are some ablations with PCA (which already are better than TIP-Adapter and TIP-X) so I was curious how well a random selection method would do?
  2. Do you have the optimal mixing hyperparameters for each dataset? I couldn't find it in the appendix.

Great work again!

yangyangyang127 commented 1 year ago

Dear Sir,

  1. We implement a random selection baseline and a no-selection baseline for APE. The 1/2/4/8/16-shot results on ImageNet under ResNet-50 backbone are shown below:

No Selection: 62.07 62.27 65.54 62.70 63.03 Random: 62.03 62.10 62.26 62.44 62.72 APE: 62.04 62.34 62.54 62.79 63.42

The random selection and no-selection approaches are both worse than the proposed criterion.

  1. The optimal hyperparameters for most datasets are given in the code except for the oxford_pets dataset, where we adjust the number of selected channels for different shot numbers, 600/700/800/900 channels can be selected.

And thanks for your attention.