🔍 About | 🚀 Quick Start | 📊 Evaluation | 🔗 Citation

🔍 About

• We propose VisionTS, a time series forecasting (TSF) foundation model building from rich, high-quality natural images 🖼️.

  • This is conceptually different from the existing TSF foundation models (text-based 📝 or time series-based 📈), but it shows a comparable or even better performance without any adaptation on time series data.

• We reformulate the TSF task as an image reconstruction task, which is further processed by a visual masked autoencoder (MAE).

🚀 Quick Start

We have uploaded our package to PyPI. Please first install pytorch, then running the following command for installing VisionTS:

pip install visionts

Then, you can refer to demo.ipynb about forecasting time series using VisionTS, with a clear visualization of the image reconstruction.

📊 Evaluation

Our repository is built on Time-Series-Library, MAE, and GluonTS. Please install the dependencies through requirements.txt before running the evaluation.

Long-Term TSF Benchmarks (Zero-Shot)

We evaluate our methods on 6 long-term TSF benchmarks for zero-shot forecasting. The scripts are under long_term_tsf/scripts/vision_ts_zeroshot. Before running, you should first follow the instructions of Time-Series-Library to download datasets into long_term_tsf/dataset. Using the following command for reproduction:

cd long_term_tsf/
bash scripts/vision_ts_zeroshot/$SOME_DATASET.sh

Monash (Zero-Shot)

We evaluate our methods on 29 Monash TSF benchmarks. You can use the following command for reproduction, where the benchmarks will be automatically downloaded.

cd eval_gluonts/
bash run_monash.sh

[!IMPORTANT] The results in the paper are evaluated based on python==3.8.18, torch==1.7.1, torchvision==0.8.2, and timm==0.3.2. Different versions may lead to slightly different performance.

PF (Zero-Shot)

We evaluate our methods on 6 long-term TSF benchmarks for zero-shot forecasting. Before running, you should first follow the instructions of Time-Series-Library to download datasets into long_term_tsf/dataset, in addition to the following three datasets:

• Walmart: https://www.kaggle.com/competitions/walmart-recruiting-store-sales-forecasting/overview (download to long_term_tsf/dataset/walmart-recruiting-store-sales-forecasting/train.csv)
• Istanbul Traffic: https://www.kaggle.com/datasets/leonardo00/istanbul-traffic-index (download to long_term_tsf/dataset/istanbul-traffic-index/istanbul_traffic.csv)
• Turkey Power: https://www.kaggle.com/datasets/dharanikra/electrical-power-demand-in-turkey (download to long_term_tsf/dataset/electrical-power-demand-in-turkey/power Generation and consumption.csv)

You can use the following command for reproduction.

cd eval_gluonts/
bash run_pf.sh

Long-Term TSF Benchmarks (Full-Shot)

We evaluate our methods on 8 long-term TSF benchmarks for full-shot forecasting. The scripts are under long_term_tsf/scripts/vision_ts_fullshot. Using the following command for reproduction:

cd long_term_tsf/
bash scripts/vision_ts_fullshot/$SOME_DATASET.sh

🔗 Citation

@misc{chen2024visionts,
      title={VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters}, 
      author={Mouxiang Chen and Lefei Shen and Zhuo Li and Xiaoyun Joy Wang and Jianling Sun and Chenghao Liu},
      year={2024},
      eprint={2408.17253},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2408.17253}, 
}

⭐ Star History