search

CVMI-Lab / SimGCD

(ICCV 2023) Parametric Classification for Generalized Category Discovery: A Baseline Study
https://arxiv.org/abs/2211.11727
MIT License
85 stars 13 forks source link

About Validation scheme #5

Closed sybapp closed 1 year ago

sybapp commented 1 year ago

I would like to thank you for your contribution. However, I have some concerns regarding the validation strategy mentioned in your paper and implemented in the code. In the paper, you mentioned that the validation strategy involves using the old class samples from the validation set as shown in the following image: image However, in the code, an unknown class selection model is used. I am wondering if this is an error or intentional. Here is the link to the relevant code: https://github.com/CVMI-Lab/SimGCD/blob/47e8374f02534a45b423f8b52e364571a551cb27/train.py#L133

xwen99 commented 1 year ago

Hi @sybapp, thanks for pointing out this issue! I have checked our development codebase, and verified that this was a mistake during the code refactoring. Please follow the description in the paper (and the original GCD codebase), and we will fix this issue soon. Thanks again!

sybapp commented 1 year ago

Is the result in the paper based on the performance of the last epoch?