BAL: Balancing Diversity and Novelty for Active Learning - Official Pytorch Implementation

Our paper has been accepted by TPAMI.

Method

Performance

Experiment Setting

Install the requirements

pip install -r requirements.txt

Prepare the dataset in the following format

- DATA_PATH
    - DATASET
        - train
            - CLS
                - *.jpg
        - test
            - CLS
                - *.jpg

e.g.

- data
    - cifar10
        - train
            - 0 
                - airplane_3.jpg
                - airplane_10.jpg
                ...
            - 1 
            ...
            - 9
        - test
            - 0
            ...
            - 9
    - caltech101
        - train
        - test
    - svhn 
        - train
        - test
    ...

Active Learning

1. To train the rotation predition task on the unlabeled set. This step will generate the SORTED_DATASET_PATH.

python rotation.py \
    --save $SAVE \
    --net vgg16 \
    --dataset cifar10 \
    --datapath $DATA_PATH \
    --lr 0.1 \
    --batch_size 256

2. To kmeans cluster pretext features and sort the unlabeled pool. LOAD_DIR refers to your pretrained weights.

python kmeans.py \
    --net vgg16 \
    --dataset cifar10 \
    --datapath $DATA_PATH \
    --load $LOAD_DIR 

3. To train and evaluate on active learning task:

python main.py \
    --net vgg16 \
    --dataset cifar10 \
    --datapath $DATA_PATH \
    --per_samples_list 10 10 10 10 10 10 10 10 10 10 \ # change it according to your AL setting
    --addendum 5000 \                                  # change it according to your AL setting
    --save $SAVE \
    --beta 1.0 \
    --milestone 30 60 90 \
    --sort high2low \
    --sampling confidence \
    --first high1st \
    --lr 0.1 \
    --sorted_dataset_path $SORTED_DATASET_PATH

Hyper-parameters

In our paper, we select the optimal ```beta`` by evaluating the results of the first epoch. You can directly utilize our experimental outcomes.

Beta	caltech101	cifar10	svhn	tinyimagenet
small	0.5	1.0	1.0	0.5
base	1.3	1.2	1.4	1.2
large	2.5	2.0	2.0	2.5

Citation

If you find our research helpful, kindly cite:

@ARTICLE{10372131,
  author={Li, Jingyao and Chen, Pengguang and Yu, Shaozuo and Liu, Shu and Jia, Jiaya},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={BAL: Balancing Diversity and Novelty for Active Learning}, 
  year={2023},
  volume={},
  number={},
  pages={1-12},
  doi={10.1109/TPAMI.2023.3345844}}

Acknowledgement

Part of the code is modified from PT4AL repo.