A Fresh Look at Generalized Category Discovery through Non-negative Matrix Factorization

Zhong Ji, Shuo Yang, Jingren Liu, Yanwei Pang, Jungong Han

---

🔥 News

• [2026-03-09] Repository initialized.
• [2026-03-09] README and project structure released.
• [2026-03-11] Training and evaluation code will be released.
• [2026-03-11] Reproducibility guide will be released.

---

Overview

Generalized Category Discovery (GCD) aims to recognize both seen (base) and unseen (novel) categories by leveraging labeled data from base classes together with unlabeled data containing both old and new classes. This repository implements NN-GCD, a simple yet effective framework that revisits GCD from a non-negative matrix factorization (NMF) perspective. The method connects:

• Optimal K-means
• Symmetric Non-negative Matrix Factorization (SNMF)
• Non-negative Contrastive Learning

and introduces tailored optimization objectives for generalized category discovery.

---

Features

• Training code for GCD benchmarks
• Single-GPU training pipeline
• Multi-GPU distributed training pipeline
• Support for standard GCD datasets
• Backbone + projector based NN-GCD implementation
• Evaluation on both old and new classes

---

Repository Structure

The current repository structure is organized as follows:

NNGCD/
├── data/                   # dataset storage / augmentation / dataset utilities
├── dataloader/             # dataset and dataloader implementations
├── models/                 # backbone definitions and model-related modules
├── scripts/                # helper scripts for launching experiments
├── util/                   # logging, evaluation, samplers, utilities
├── config.py               # global config (e.g. experiment root)
├── model.py                # core loss functions / heads / view generators / helpers
├── train.py                # single-GPU training entry
├── train_mp.py             # multi-GPU distributed training entry
└── requirements.txt         # dependencies

---

Installation

1. Create environment

conda create -n nngcd python=3.10 -y
conda activate nngcd

2. Install dependencies

pip install -r requirements.txt

---

Datasets

This repository supports common GCD benchmarks, including: CIFAR-10, CIFAR-100, ImageNet-100, CUB, Stanford Cars, FGVC Aircraft and Herbarium-19. Please download the datasets and place them under the data/ directory.

A recommended structure is:

data/
├── cifar/
├── cub/
├── fgvc_aircraft/
├── herbarium_19/
├── imagenet/
├── stanford_cars/
└── ssb_splits/

---

Training

This project provides two training modes:

1. Single-GPU / standard training using train.py
2. Multi-GPU distributed training using train_mp.py

Use train.py when you want to run experiments on a single GPU.

Example

python train.py \
  --dataset_name cifar10 \
  --batch_size 128 \
  --epochs 200 \
  --lr 0.1 \
  --momentum 0.9 \
  --weight_decay 1e-4 \
  --transform imagenet \
  --sup_weight 0.35 \
  --reg_weight 3e-5 \
  --sparsity_weight 0.6 \
  --nce_temp 0.5 \
  --n_views 2 \
  --memax_weight 2 \
  --warmup_teacher_temp 0.07 \
  --teacher_temp 0.04 \
  --warmup_teacher_temp_epochs 30 \
  --student_temp 0.1 \
  --fp16 \
  --print_freq 10 \
  --exp_name nngcd_cifar10_single

Use train_mp.py for distributed training across multiple GPUs.

Example (4 GPUs on one machine)

torchrun --nproc_per_node=4 train_mp.py \
  --dataset_name cifar10 \
  --batch_size 128 \
  --epochs 200 \
  --lr 0.1 \
  --momentum 0.9 \
  --weight_decay 1e-4 \
  --transform imagenet \
  --sup_weight 0.35 \
  --n_views 2 \
  --memax_weight 2 \
  --warmup_teacher_temp 0.07 \
  --teacher_temp 0.04 \
  --warmup_teacher_temp_epochs 30 \
  --fp16 \
  --print_freq 10 \
  --exp_name nngcd_cifar10_ddp

A pretrained or previously saved checkpoint can be loaded using --warmup_model_dir.

Example

python train.py \
  --dataset_name cifar10 \
  --warmup_model_dir path/to/checkpoint.pt \
  --exp_name eval_or_finetune_run

or for distributed mode:

torchrun --nproc_per_node=4 train_mp.py \
  --dataset_name cifar10 \
  --warmup_model_dir path/to/checkpoint.pt \
  --exp_name eval_or_finetune_run_ddp

---

Citation

If you find this repository useful, please cite:

@article{ji2026fresh,
  title={A fresh look at generalized category discovery through non-negative matrix factorization},
  author={Ji, Zhong and Yang, Shuo and Liu, Jingren and Pang, Yanwei and Tang, Chen},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  publisher={IEEE}
}

---

⭐ Star History

---

📬 Contact

For questions and discussions:

• Email: jrl0219@tju.edu.cn
• Issues: please open a GitHub Issue in this repo.

---