Giters

wangkiw / ICLR23-MEMO

The code repository for "A Model or 603 Exemplars: Towards Memory-Efficient Class-Incremental Learning" (ICLR'23) in PyTorch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Some confusion about parameters and loss functions

vanity1129 opened this issue · comments

commented

Thanks for your excellent work!

But I have some confusion about the params "-ms 3312", can you explain why you chose this parameter size?
And about the loss function "loss_aux=F.cross_entropy(aux_logits,aux_targets)", I noted the definition of the FC layer, "self.aux_fc=self.generate_fc(self.out_dim, new_task_size+1)", can you provide more explanation about aux_fc?

Many thanks and looking forward to the reply

commented

Hi, Thank you for your attention.

  • The concrete memory size we use can refer to the supplementary material of our paper.
  • For the loss function, we inherit it from DER. It takes all old classes as an auxiliary class.
commented

Thanks for your reply!