DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025)

DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery
By Yuanpei Liu and Kai Han.

Prerequisite 🛠️

First, you need to clone the DebGCD repository from GitHub. Open the terminal and run the following command:

git clone https://github.com/Visual-AI/DebGCD.git
cd DebGCD

We recommend setting up a conda environment for the project:

conda create --name=debgcd python=3.8
conda activate debgcd
pip install -r requirements.txt

Running 🏃

Config

Set paths to datasets, pretrained weights, and log directories in config.py

Datasets

We use generic object recognition datasets, including CIFAR-10/100 and ImageNet-100:

• CIFAR-10/100 and ImageNet-100

We also use fine-grained benchmarks (CUB, Stanford-cars, FGVC-aircraft). You can find the datasets in:

• The Semantic Shift Benchmark (SSB)

Scripts

We use the slurm system to run the code. The scripts to train and eval DebGCD models on different datasets can be found in the folder /scripts. For example, to train and eval on Stanford Cars dataset.

Eval the model

sbatch scripts/eval_DebGCD.cmd scars

Train the model:

sbatch scripts/train_scars.cmd

Please note that we have further tuned the hyperparameters to get optimal performance on each dataset, which can be slightly different under different conda environments. So, it's suggested to use install the environment following the provided requirements. Our models can be downloaded from this link.

Citing this work

If you find this repo useful for your research, please consider citing our paper:

@inproceedings{liu2025debgcd,
  title={DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery},
  author={Liu, Yuanpei and Han, Kai},
  booktitle={ICLR},
  year={2025}
}