I have multiple DNN that can extract features vector from images. Those can be used for two main goals:

  1. Use them for transfer learning ang faster trainings
  2. Use them as feature extractors, and train only additional, small "heads" (in particular if my dataset is really small)

But how can I choose which trained network to use? For the moment, I use:

  • A UMAP projection on a labelled dataset (that the networks have never seen), and a compare visually how well my various classes are "clustered". enter image description here
  • A "Raw 10-NN Mean" score. For each point, I count how many of their 10 nearest neighbours are of the same class. I Divide by 10 to get the score.
  • A "Class 10-NN Mean" score. Same as before, but each class has the same weight, regardless of their proportion in the dataset.

Are there any other interesting metrics I could use to compare my feature extractors?

Many thanks!



You must log in to answer this question.

Browse other questions tagged .