Neighborhood Contrastive Learning for Novel Class Discovery

This repository contains the official implementation of our paper:

Neighborhood Contrastive Learning for Novel Class Discovery, CVPR 2021
Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, Nicu Sebe

Requirements

PyTorch >= 1.1

Data preparation

We follow AutoNovel to prepare the data

By default, we save the dataset in ./data/datasets/ and trained models in ./data/experiments/.

For CIFAR-10 and CIFAR-100, the datasets can be automatically downloaded by PyTorch.
For ImageNet, we use the exact split files used in the experiments following existing work. To download the split files, run the command: sh scripts/download_imagenet_splits.sh . The ImageNet dataset folder is organized in the following way:
```
ImageNet/imagenet_rand118 #downloaded by the above command
ImageNet/images/train #standard ImageNet training split
ImageNet/images/val #standard ImageNet validation split
```

Pretrained models

We use the pretrained models (self-supervised learning and supervised learning) provided by AutoNovel. To download, run:

sh scripts/download_pretrained_models.sh

If you would like to train the self-supervised learning and supervised learning models by yourself, please refer to AutoNovel for more details.

After downloading, you can go to perform our neighbor contrastive learning below.

Neighborhood Contrastive Learning for Novel Class Discovery

CIFAR10/CIFAR100

Without Hard Negative Generation (w/o HNG)

# Train on CIFAR10
CUDA_VISIBLE_DEVICES=0 sh scripts/ncl_cifar10.sh ./data/datasets/CIFAR/ ./data/experiments/ ./data/experiments/pretrained/supervised_learning/resnet_rotnet_cifar10.pth

# Train on CIFAR100
CUDA_VISIBLE_DEVICES=0 sh scripts/ncl_cifar100.sh ./data/datasets/CIFAR/ ./data/experiments/ ./data/experiments/pretrained/supervised_learning/resnet_rotnet_cifar100.pth

With Hard Negative Generation (w/ HNG)

# Train on CIFAR10
CUDA_VISIBLE_DEVICES=0 sh scripts/ncl_hng_cifar10.sh ./data/datasets/CIFAR/ ./data/experiments/ ./data/experiments/pretrained/supervised_learning/resnet_rotnet_cifar10.pth

# Train on CIFAR100
CUDA_VISIBLE_DEVICES=0 sh scripts/ncl_hng_cifar100.sh ./data/datasets/CIFAR/ ./data/experiments/ ./data/experiments/pretrained/supervised_learning/resnet_rotnet_cifar100.pth

Note that, for cifar-10, we suggest to train the model w/o HNG, because the results of w HNG and w/o HNG for cifar-10 are similar. In addition, the model w/ HNG sometimes will collapse, but you can try different seeds to get the normal result.

ImageNet

Without Hard Negative Generation (w/o HNG)

# Subset A
CUDA_VISIBLE_DEVICES=0 python ncl_imagenet.py --unlabeled_subset A --model_name resnet_imagenet_ncl

# Subset B
CUDA_VISIBLE_DEVICES=0 python ncl_imagenet.py --unlabeled_subset B --model_name resnet_imagenet_ncl

# Subset C
CUDA_VISIBLE_DEVICES=0 python ncl_imagenet.py --unlabeled_subset C --model_name resnet_imagenet_ncl

With Hard Negative Generation (w/o HNG)

# Subset A
CUDA_VISIBLE_DEVICES=0 python ncl_imagenet.py --hard_negative_start 3 --unlabeled_subset A --model_name resnet_imagenet_ncl_hng

# Subset B
CUDA_VISIBLE_DEVICES=0 python ncl_imagenet.py --hard_negative_start 3 --unlabeled_subset B --model_name resnet_imagenet_ncl_hng

# Subset C
CUDA_VISIBLE_DEVICES=0 python ncl_imagenet.py --hard_negative_start 3 --unlabeled_subset C --model_name resnet_imagenet_ncl_hng

Acknowledgement

Our code is heavily designed based on AutoNovel. If you use this code, please also acknowledge their paper.

Citation

We hope you find our work useful. If you would like to acknowledge it in your project, please use the following citation:

@InProceedings{Zhong_2021_CVPR,
      author    = {Zhong, Zhun and Fini, Enrico and Roy, Subhankar and Luo, Zhiming and Ricci, Elisa and Sebe, Nicu},
      title     = {Neighborhood Contrastive Learning for Novel Class Discovery},
      booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      month     = {June},
      year      = {2021},
      pages     = {10867-10875}
}

Contact me

If you have any questions about this code, please do not hesitate to contact me.