Contrastive Subspace Distribution Learning for Novel Category Discovery in High-Dimensional Visual Data
This repository contians the Pytorch implementation of our paper Contrastive Subspace Distribution Learning for Novel Category Discovery in High-Dimensional Visual Data
pip install -r requirements.txt
sh train.sh
Configure the dataset path according to
config.py
Our main results accelarate on generic image recognition datasets
| Dataset | All | Old | New |
|---|---|---|---|
| CIFAR10 | 0.9666 | 0.9696 | 0.9696 |
| Fashion-Mnist | 0.8725 | 0.9466 | 0.8355 |
| EMNIST | 0.7804 | 0.7933 | 0.6865 |
| CIFAR100 | 0.7857 | 0.7818 | 0.7929 |
The unsupervised and semi-supervised results accelarate
| Dataset | unsupervised | semi-supervised |
|---|---|---|
| CIFAR10 | 0.8768 | 0.9673 |
| Fashion-Mnist | 0.6630 | 0.9155 |
| EMNIST | 0.7030 | 0.9502 |
| CIFAR100 | 0.5745 | 0.8289 |
This project is licensed under the MIT License - see the LICENSE file for details.
If you find this repository useful, please consider giving a citation