Skip to content

Regarding Usage of EWC strategy and Replay strategy #1065

Answered by AndreaCossu
AdityaJainTD asked this question in Q&A
Discussion options

You must be logged in to vote
  1. EWC automatically computes importances. Importance-based regularization strategies are well-known to not work in New Classes scenario so your results are correct.
  2. Replay keeps a fixed memory (of 200 previous samples in your cases) and balances it across the encountered classes. So each time, you will have in your memory 200/n_classes_encountered_so_far per examples per class in your memory. Replay concatenates to the current minibatch a random sample from the memory and trains the model on this extended minibatch.

Hope this helps 😄

Replies: 1 comment 6 replies

Comment options

You must be logged in to vote
6 replies
@AndreaCossu
Comment options

@AntonioCarta
Comment options

@AdityaJainTD
Comment options

@AndreaCossu
Comment options

@AdityaJainTD
Comment options

Answer selected by AdityaJainTD
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Category
Q&A
Labels
None yet
3 participants