A Modular Framework for Multi-step Class-incremental Novel Class Discovery

Intro • Key Features • License • Cite

Large-scale Pre-trained Models are Surprisingly Strong in Incremental Novel Class Discovery

Mingxuan Liu, Subhankar Roy, Zhun Zhong, Nicu Sebe, and Elisa Ricci

This Github repository first presents the PyTorch implementation for the paper (pre-printed) Large-scale Pre-trained Models are Surprisingly Strong in Incremental Novel Class Discovery [arXiv].

Code coming soon ☘️

Introduction

A flexible modular PyTorch repository for the MSc-iNCD task, allowing for easy replication and adoption of not only our baselines, but also typical Incremental Learning (IL), Novel Class Discovery (NCD), and class-incremental Novel Class Discovery (class-iNCD) methods that adapted to the new setting. It supports representative NCD (RankStats, OCRA, Sinkhorn+Swap) and IL (EwC, LwF, DER, CosNorm) components that allow users to develop new methods for the MSc-iNCD task.

Key Features

We provide a flexible modular framework that supports representative NCD (RankStats, OCRA, Sinkhorn+Swap) and IL (EwC, LwF, DER, CosNorm) components, allowing users to develop new methods for the MSc-iNCD task. Experiments by default provide results on MSc-iNCD setting starting from a large-scale pre-trained model. Furthermore, it supports supervised pre-training (used in NCD and class-iNCD) by using a certain (user-defined) amount of the novel classes as base classes in each data set.

Setting	Supervised Pre-train	Large-scale Pre-train	Data Access
Novel Class Discovery	✅	❌	✅
Class-incremental Novel Class Discovery	✅	❌	❌
Multi-step Class-incremental Novel Class Discovery	❌	✅	❌

Current available approaches and schemes include:

Training Scheme: Fine-tuning • Freezing • Joint

Incremental components: EwC • LwF • DER • FRoST • ResTune • MSc-iNCD (Ours)

Discovery components: RankStats • OCRA • Sinkhorn+Swap (Ours)

License

Please check the MIT license that is listed in this repository.

Cite

If you find our framework or paper useful, please cite:

@article{Liu2023LargescalePM,
  title={Large-scale Pre-trained Models are Surprisingly Strong in Incremental Novel Class Discovery},
  author={Mingxuan Liu and Subhankar Roy and Zhun Zhong and Nicu Sebe and Elisa Ricci},
  journal={arXiv preprint arXiv:2303.15975},
  year={2023}
}

The codebase for the adapted methods is created by FRoST, ResTune, Mammoth, AutoNovel, and OCRA. If you find these adapted methods useful, it would be appreciated if you acknowledge the original papers by citing them using the name and URL mentioned before.