Learn to Categorize or Categorize to Learn? Self-Coding for Generalized Category Discovery (NeurIPS 2023)
By
Sarah Rastegar,
Hazel Doughty, and
Cees Snoek.
pip install -r requirements.txt
Since our work relies heavily on kmeans_pytorch for cluster assignments, you need to ensure that it is correctly imported to reproduce the results from the paper. You can install kmeans_pytorch directly in the directory by executing the following commands:
cd InfoSieve
git clone https://github.com/subhadarship/kmeans_pytorch
cd kmeans_pytorch
pip install --editable .
Note: While using scikit-learn's KMeans provides improvements, the results in the paper have been reported using kmeans_pytorch.
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.
We use fine-grained benchmarks in this paper, including:
We also use generic object recognition datasets, including:
- CIFAR-10/100 and ImageNet
Train representation: To run the code with the hyperparameters used in the paper, execute the following command:
python contrastive_training.py
This script will automatically train the representations, extract features, and fit the semi-supervised KMeans algorithm. It also provides final evaluations on both the best checkpoint and the final checkpoint.
Dataset Hyperparameter Specifics: If you're working with the CUB dataset, set the unsupervised_smoothing parameter to 1.0, for other fine-grained datasets (Scars, Pets, Aircraft) to 0.5 and for generic datasets to 0.1. Please note that for Herbarium_19, --unbalanced 1 should be added, because it is long-tailed.
In the Final Reports section at the end, please note that only the evaluations reported for:
Reports for the best checkpoint:
Reports for the last checkpoint:
are the evaluations performed at test time to evaluate the checkpoints. Additionally, please note that Train ACC Unlabelled_v2 is the metric reported by our work and prior studies.
If you use this code in your research, please consider citing our paper:
@inproceedings{
rastegar2023learn,
title={Learn to Categorize or Categorize to Learn? Self-Coding for Generalized Category Discovery},
author={Sarah Rastegar and Hazel Doughty and Cees Snoek},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=m0vfXMrLwF}
}
The codebase is mainly built on the repo of https://github.com/sgvaze/generalized-category-discovery.
If you found our code helpful and are interested in exploring more, also check out our ECCV 2024 paper SelEx: Self-Expertise in Fine-Grained Generalized Category Discovery 
