Labels for task>0

[Areeb2735]

Labels for task>0

[zhoudw-zdw]

Hi @Areeb2735, thanks for your interest in our work.

The classes are not mapped to [0,1,2,3,4,5,6,7,8,9], and they are actually in the format of [10,11,12,13,14,15,16,17,18,19]. You can refer to this function for more details. https://github.com/zhoudw-zdw/CIL_Survey/blob/791728c072974da264d985d7e9c4825fdc3d17ff/utils/data_manager.py#L137C5-L137C5

Additionally, you can print the labels during training to see the labels.

[Areeb2735]

Thanks.

But then how is the loss computed? Given that you are using cross-entropy loss, the targets are the index. Now the targets are [10,11,12,13,14,15,16,17,18,19].

[zhoudw-zdw]

In class-incremental learning, the classifier is expanded after each task.

So you have a classifier that outputs 20 dims at the second stage, 30 for the third stage, etc.

See here for the classifier update.

CIL_Survey/models/finetune.py

Line 43 in 791728c

self._network.update_fc(self._total_classes)

[Areeb2735]

Thank you