search

mmasana / FACIL

Framework for Analysis of Class-Incremental Learning with 12 state-of-the-art methods and 3 baselines.
https://arxiv.org/pdf/2010.15277.pdf
MIT License
512 stars 98 forks source link

Some question about the approach 'BiC' #16

Closed ZhenbangDu closed 1 year ago

ZhenbangDu commented 2 years ago

Hello! Thank you for your amazing work! But it seems that in BiC, in line 102, when creating the dataset val_old, the code doesn't consider the situation that "self.exemplars_dataset.max_num_exemplars_per_class != 0".

mmasana commented 2 years ago

Hi @YixianDu, happy to hear you like our framework!

But it seems that in BiC, in line 102, when creating the dataset val_old, the code doesn't consider the situation that "self.exemplars_dataset.max_num_exemplars_per_class != 0".

That is correct. In the condition of line 102 we remove extra exemplars from previous classes. But this is only done when the type of memory is fixed (and therefore has a limit). In the case of a growing memory, we do not want to remove any exemplar, and that's why we do not check if self.exemplars_dataset.max_num_exemplars_per_class != 0.

Hope that solves the confusion :)