Awesome

Time Series Dataset Condensation via Dual Domain Matching [KDD2024]

Data

The data is in https://www.dropbox.com/scl/fi/izu5d5gto7q84tezc2pc1/TimeSeriesData.zip?rlkey=qeqd2qeyse6trlm1t6yf6bppu&dl=0

Please download it and unzip it in ./data folder

Environment

The code is implemented in pytorch 1.10.0, CUDA version 11.3, python 3.7.0.

pip install torch==1.10.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

Reproducibility

The default configs of the four datasets are set in ./config. To reproduce the result, first we need to get the training trajectory:

bash DCDDM_buffer.sh

Then we need to run the training code

bash DCDDM_distill.sh

The default hyper-parameters and experiment setting is in the .sh file. Default we run the experiment for HAR dataset of $spc=5$.