Contrastive Learning for Practical Synthetic Data Generation Using Seasonal and Trend Representations
- Download the datasets and place them in
CoST-main/datasets - For training, run
pip install -r requirements.txton terminal followed by:- mendeley :
python -u train.py mendeley forecast_multivar --alpha 0.0005 --kernels 1 2 4 8 16 32 64 128 --max-train-length 201 --batch-size 128 --archive forecast_csv --repr-dims 320 --max-threads 8 --seed 2 --eval --epochs 400 - calce :
python -u train.py dataset4 forecast_dataset4 --alpha 0.0005 --kernels 1 2 4 8 16 32 64 128 --max-train-length 201 --batch-size 128 --archive forecast_csv_univar --repr-dims 320 --max-threads 8 --eval --epochs 400 - colab implementation
- mendeley :
- After training is completed, the results are stored in
CoST-main/training - The results can be viwed by running
We would like to thank the contributors of the original CoST repository and the authors of the paper titled CONTRASTIVE LEARNING OF DISENTANGLED SEASONAL-TREND REPRESENTATIONS FOR TIME SERIES FORECASTING
Additionally we would like to mention that the being used is sourced from:
@inproceedings{
woo2022cost,
title={Co{ST}: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting},
author={Gerald Woo and Chenghao Liu and Doyen Sahoo and Akshat Kumar and Steven Hoi},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=PilZY3omXV2}
}