Leveraging 2D Information for Long-term Time Series Forecasting with Vanilla Transformers

A simple yet strong Long-term Time Series prediction model.

We recommand to use Conda to mange a virtual environment:

conda create -n gridtst python=3.8 && conda activate gridtst
pip install -r requirements.txt

logging and multi-gpu training setup:

wandb login
accelerate config

This is the dataset we use, you could download here and put all csv files in the dataset folder.

We provide all the scripts on the scripts folder. For example, training on the Weather dataset with lookback window = 336:

bash scripts/lookback_window_336/weather.sh

We provide our trained model on the huggingface space

To evaluate these models, you could either specify a perticular model or evaluate them all at once.

For a certain model, for example GridTST on traffic dataset with lookback window=336 and prediction length=96:

python benchmark.py --data_file dataset/traffic.csv --seq_len 336 --label_len 96

To evaluate them all:

python benchmark.py --all