This github is the implementation of ICLR'22 paper, named as Coherence-based Label Propagation over Time Series for Accelerated Active Learning. Please follow the instructions to reproduce our work.
pip install -r requirements.txtDownload I3D feature data of 50salads and GTEA at link and locate the contents at the ./datasets/DATASET_NAME. The link is from ms-tcn, the repository for the paper "Y. Abu Farha and J. Gall. MS-TCN: Multi-Stage Temporal Convolutional Network for Action Segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019."
For HAPT and mHealth dataset, download HAPT and mHealth and locate the contents at the ./datasets/DATASET_NAME
At current directory which has all source codes, run main.py with the parameters as follows.
- dataset: {50salads, GTEA, mHealth, HAPT} designates which dataset to use.
- seed: {0, 1, 2, 3, 4} is the seed for 5-fold cross validation.
- gpu: {0, 1, 2, ...} is an integer for gpu id
- lp: {platprob, repr, prob, zero} indicates label propagation method to use, representing {TCLP, ESP, PTP, NOP} in the paper, respectively.
- al: {conf, entropy, margin, core, badge, utility} shows active learning to use, representing {CONF, ENTROPY, MARG, CS, BADGE, UTILITY} in the paper, respectively.
- no_plat_reg: {0, 1} decides whether or not to use width regularization or not. 1 means removing width regularization.
- temp: [1, infinity] is the parameter T for temperature scaling. T=1 means no temperature scaling.
Here's the example running code.
python3 main.py --dataset HAPT --gpu 0 --seed 0 --lp platprob --al random --no_plat_reg 1 --temp 2.0- Classification accuracy at each active learning round is saved in
metadatafolder as.npyformat.