Towards On-the-Fly Novel Category Discovery in Dynamic Long-Tailed Distributions

Introduction

NCD-DLT aims to discover novel classes in a challenging condition, where new unlabeled data keeps arriving dynamically within the long-tailed distribution.

Install

Requirements

cuda=11.6.1
numpy=1.23.5
pandas=1.5.3
python=3.10.9
torch=2.0.1

Clone Repository

git clone https://github.com/HoinJung/NCD-DLT.git

Pretrained DINO

1. Download the ViT-B/16 backbone checkpoint from this repository. Link
2. Locate the DINO's pretrained weight 'pretrained_models' directory.

Run code

• --greedy: Use greedy hash regulazation loss.
• --double: Use double hashing in the projection head.
• --pseudo: Use the Hamming hash graph merging algorithm as a pseudo label with supervised contrastive learning.
• --threshold_size: Node size threshold for graph merging.
• --threshold_confidence: Confidence threshold for graph mering.
• --lc or --no-lc: Long-tailed distribution or not.
• --imb_ratio: Imbalance ratio of the long-tailed distribution.
• Post-processing with graph merging will be applied automatically. Both the naive accuracy and post-processed accuracy will displayed together.

python NCD-DLT.py --gpu_id 0  --init_epochs 50 --epochs 10  --dataset_name cifar100 --greedy --double --pseudo --lam 10 --threshold_size 10 --threshold_confidence 0.2 --lc --imb_ratio 100

Test

During the training, the evaluation is conducted at every epoch in the initial stage and the last epoch of every dynamic stage.

Citation

TBD

LICENSE

MIT License