HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting

Installation

1. Dataset Installation

Run the following command to download and preprocess the dataset (Taxi as an example) in each training and inference stage:

dataset = get_dataset("taxi_30min", regenerate=True)  # Set regenerate=True for the first time

Training

Stage 1 Training

python ./stage1_downsampled_target/stage1_dowsample_run.py  # Discrete downsampled target training
python ./stage1_target/stage1_run.py   # Discrete target training

Stage 2 Training

python ./Stage2_downsampled_generation/main.py  # Discrete downsampled target generation
python ./Stage2_target_generation/main.py   # Discrete target generation

Eval

Inference

python ./Stage2_inference/inference.py  # Target forecasting

Example Checkpoint

https://drive.google.com/file/d/18_cDNGP8yB8AT48IARivGfXzMMczJ6E-/view?usp=sharing

Citing

To cite this repository:

@software{pytorchgithub,
    author = {Kashif Rasul},
    title = {{P}yTorch{TS}},
    url = {https://github.com/zalandoresearch/pytorch-ts},
    version = {0.6.x},
    year = {2021},
}

@article{feng2025hdt,
  title={HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting},
  author={Feng, Shibo and Zhao, Peilin and Liu, Liu and Wu, Pengcheng and Shen, Zhiqi},
  journal={arXiv preprint arXiv:2502.08302},
  year={2025}
}