Structured Component-based Neural Network

This is an implementation of SCNN.

Requirements

Python 3.7
numpy >= 1.18.5
pandas >= 1.0.3
torch >= 1.12.0
pytorchts == 0.6.0
h5py
pytorchts

Model Training

python -u main.py -mode train -short_term 8 -long_term 144 -n_local_input 2 -cuda 0

Arguments

short_term: length of short term.
long_term: length of long term.
dataset: dataset name.
version: version number.
hidden_channels: number of hidden channels.
n_pred: number of output steps.
n_his: number of input steps.
n_local_input: kernel size of causual convolution.
n_layers: number of hidden layers.
cuda: cuda device id.

Model Evaluation

python -u main.py -mode eval -cuda 0

Citation

@misc{deng2023learning,
      title={Learning Structured Components: Towards Modular and Interpretable Multivariate Time Series Forecasting}, 
      author={Jinliang Deng and Xiusi Chen and Renhe Jiang and Du Yin and Yi Yang and Xuan Song and Ivor W. Tsang},
      year={2023},
      eprint={2305.13036},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

jiawei0322 / SCNN