NeuralCollapseApplications / FSCIL

[ICLR 2023] The official code for our ICLR 2023 (top25%) paper: "Neural Collapse Inspired Feature-Classifier Alignment for Few-Shot Class-Incremental Learning"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

produce training rect

gulzainali98 opened this issue · comments

I am not sure what produce_training_rect function is supposed to do and why is it used for few shot sessions only?

It seems it is weighting weights of classifier up and down based on the number of samples in a mini batch.

yes, in incremental sessions, train samples of novel classes and features of old classes from the memory are imbalanced, so we use this re-weight strategy. Base session has balanced training data.