Generalized Category Discovery

This repo contains code for our paper: Generalized Category Discovery

Given a dataset, some of which is labelled, Generalized Category Discovery is the task of assigning a category to all the unlabelled instances. Unlabelled instances could come from labelled or 'New' classes.

Contents

💥 1. Updates

🏃 2. Running

🔢 3. Results

📋 4. Citation

💥 Updates

Updates to paper since pre-print (updated PDF available here, ArXiv updating soon)

• We introduced a more rigorous evaluation metric - when computing ACC, we compute the Hungarian algorithm only once across all unlabelled data.
  • This single set of linear assignments is then used to compute ACC on 'Old' and 'New' class subsets (see Appendix E)
  • Practically, this involves switching from 'v1' to 'v2' evaluation in ./project_utils/cluster_and_log_utils.py

🏃 Running

Dependencies

pip install -r requirements.txt

Config

Set paths to datasets, pre-trained models and desired log directories in config.py

Set SAVE_DIR (logfile destination) and PYTHON (path to python interpreter) in bash_scripts scripts.

Datasets

We use fine-grained benchmarks in this paper, including:

• The Semantic Shift Benchmark (SSB) and Herbarium19

We also use generic object recognition datasets, including:

• CIFAR-10/100 and ImageNet

Scripts

Train representation:

bash bash_scripts/contrastive_train.sh

Extract features: Extract features to prepare for semi-supervised k-means. It will require changing the path for the model with which to extract features in warmup_model_dir

bash bash_scripts/extract_features.sh

Fit semi-supervised k-means:

bash bash_scripts/k_means.sh

Note on semi-supervised k-means

Under the old evaluation metric ('v1') we found that semi-supervised k-means consistently boosted performance over standard k-means, on 'Old' and 'New' data subsets. When we changed to 'v2' evaluation, we re-evaluated models in Tables {2,3,5} (including the ablation) and updated the figures.

However, recently, we have found that SS-k-means can be sensitive to bad initialisation under 'v2', and can sometimes lower performance on some datasets. Increasing the number of inits for SS-k-means can help. We are investigating this further now - suggestions and PRs welcome!

🔢 Results

Results from re-running models with this repo compared to reported numbers:

Dataset	All	Old	New
Stanford Cars (paper)	39.0	57.6	29.9
Stanford Cars (repo)	39.9	58.5	30.9
CIFAR100 (paper)	70.8	77.6	57.0
CIFAR100 (repo)	71.3	77.4	59.1

📋 Citation

If you use this code in your research, please consider citing our paper:

@InProceedings{vaze2022gcd,
               title={Generalized Category Discovery},
               author={Sagar Vaze and Kai Han and Andrea Vedaldi and Andrew Zisserman},
               booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
               year={2022}}