search

Impression2805 / CVPR21_PASS

PyTorch implementation of our CVPR2021 (oral) paper "Prototype Augmentation and Self-Supervision for Incremental Learning"
104 stars 20 forks source link

NCM for classification #3

Closed TamashaM closed 3 years ago

TamashaM commented 3 years ago

First of all, thank you for the great code repository. I noticed that you have mentioned in the paper that you are using the nearest class mean (NCM) classifier to classify between the classes. However, I could not locate the NCM classifier within the code. Could you please clarify?

ChibisukeDragon commented 2 years ago

@TamashaM I have the same confusion about this issue. Have you found the code of NCM?

TamashaM commented 2 years ago

@ChibisukeDragon I contacted the authors and found out that the nearest class mean (NCM) classifier is used only for the preliminary experiments in Section 3.5.2: “A Closer Look at SSL for CIL”. They further explained that the proof-of-concept experiments in Section 3.5.2 aim to demonstrate the effectiveness of self-supervised learning (SSL) for representation learning. Hence, they use the NCM classifier, which could reflect the discriminability of feature representations. While for the experiments in Section 4, they use the model outputs, i.e., the original linear classifier of the model.

ChibisukeDragon commented 2 years ago

@TamashaM Thanks a lot! You did a great job!