The confusion about the definition of FSCIL in your paper and code

[Arlannn1]

Hello, first thank you for sharing the code! It was so glad to see your works.

But I have some questions about the definition of FSCIL in this paper and code.

As we all known, in 2013 Goodfellow posted catastrophic forgetting problem，because the novel seesions have no access to the data of the previous sessions.

In the 22nd quotation in your paper, 2020 CVPR《Few-Shot Class-Incremental Learning》, this paper put forward FSCIL for the first time, and the definition of FSCIL in this paper is that , is the set of classes of the t-th training set. is the few-shot training set of new classes and only is available at the 𝑡-th training session.

But in your paper, your definition of FSCIL is that in 𝑡-th streaming session is novel classes, . And you denote the totally encountered class in 𝑡-th session as .

Obviously, you have different definitions.

From this point of view, at the 𝑡-th training session, the sample set in your definition is large, including the previous sessions, which is inconsistent with the setting of FSCIL proposed for the first time. After reading your code, I found that you really wrote the code according to your definition. That is to say, in each session, the support set is large, including all the classes seen before, which may also be inconsistent with the setting of few-shot learning. In addition, when debugging, I also found that the indexes of the support set and the query set overlap, that is https://github.com/RobinLu1209/Geometer/blob/237aed033acd30af67cbd07b430dd20dcf3403be/main.py#L270-L272

I hope I can get your help and look forward to your reply very sincerely！