Training Ensembles with Inliers and Outliers for Semi-supervised Active Learning

This repository contains the code, data, and models from the paper Vladan Stojnić, Zakaria Laskar, Giorgos Tolias, "Training Ensembles with Inliers and Outliers for Semi-supervised Active Learning", In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024.

Setup

This code was implemented using Python 3.10., to install the required dependancies please use supplied requirements.txt file.

Data

Data and splits for labeled and unlabeled datasets is available here. Please download them and put them in data directory of this repository.

Splits for initial labeled and unlabeled sets are given in splits directory, pickle files given there contain dictionary with a structure:

- seed {0, 1, 2, 3, 4}
    - train
        - labeled
            - k_shot {1, 5, 20} # list of images in the initial labeled set, for different number of images per class
        - unlabeled
            - outlier_ratio {0, 5, 20, 50, 80, 90} # list of images in the unlabeled set for specified outlier ratio
    - val # list of images for the test set, this list is same for all seeds

Models

Pre-trained self-supervised models used as initialization for all active learning methods are available here. Please download them and put them in models/ssl directory of this repository.

Running the code

To run our methods please use commands as follows:

• For the variant of our method with semi-supervision

python src/main.py dataset=imagenet seed=0 k_shot=20 al_rounds=11 use_outlier_class=true ensemble_len=5 budget=500 strategy=vr outlier_ratio=50 use_filtering=true use_semi=true epochs_semi=3

to run for other seed specify seed to a value from {0,1,2,3,4} and to run on a different outlier ratio specify outlier_ratio to a value from {0,5,20,50,80,90}.

• For the variant of our method without semi-supervision

python src/main.py dataset=imagenet seed=0 k_shot=20 al_rounds=11 use_outlier_class=true ensemble_len=5 budget=500 strategy=vr outlier_ratio=50 use_filtering=true use_semi=false

you can use the same values for seed and outlier_ratio as in the variant with semi-supervision.

To run on other datasets change dataset to cifar100, tiny_imagenet or imagenet_carnivore. For cifar100 and tiny_imagenet only outlier_ratio of 80 is available, while for imagenet_carnivore only outlier_ratio of 50 is available. Data splits are availble for different outlier ratios as well, but pre-trained self-supervised models are not.

For imagenet and imagenet_carnivore in conf/dataset/imagenet.yaml and conf/dataset/imagenet_carnivore.yaml set full_imagenet_path to the path of downloaded imagenet dataset.

Other active learning methods from the paper coming soon.

Citation

@InProceedings{Stojnic_2024_WACV,
    author    = {Stojni\'c, Vladan and Laskar, Zakaria and Tolias, Giorgos},
    title     = {Training Ensembles With Inliers and Outliers for Semi-Supervised Active Learning},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {260-269}
}