This repository contains the code of AEC-GAN: Adversarial Error Correction GANs for Auto-Regressive Long Time-Series Generation.

Getting Started:

1. Prepare the data:
   • We have provide the data in the folder data, containing the following six datasets:
     • [etth1, etth2, ettm1, ettm2, us_births, ILI]
2. Install dependencies:
   • This project is implemented with pytorch==1.8.1+cu102
3. For training

   python train.py -datasets $dataset -base_dir $save_path -p $p -q $q -use_cuda -algos 'AECGAN' -total_steps 10000 -batch_size 200 -noise_type min_adv -use_ec 2 

   • $dataset: We have implemented AEC-GAN on six datasets: [etth1, etth2, ettm1, ettm2, us_births, ILI]
   • $save_path: The path you save the model.
   • $p: The length of the past conditions.
   • $q: The length of the forward generations.
4. For generation

   python train.py -datasets $dataset -base_dir 'results/p168_q336' -p 168 -q 336 -use_cuda -algos 'AECGAN' -total_steps 10000 -batch_size 200 -noise_type min_adv -use_ec 2 -test 

   • $dataset: We have implemented AEC-GAN on six datasets: [etth1, etth2, ettm1, ettm2, us_births, ILI]
   • $save_path: The path you save the model.
   • $p: The length of the past conditions.
   • $q: The length of the forward generations.
5. For downstream performance
   • The generated data can be used as the alternative training set used for training forecasting models (e.g., SCINet, Informer and Autoformer).
   • Folder models contains the code to train the downstream forecasting models.
6. Easy usage
   • For an easy usage, we also provide a bash file run_file.sh, which contains the commands to train the models or generate time-series data.

Resources

• The GANs' code is partially based on the https://github.com/SigCGANs/Conditional-Sig-Wasserstein-GANs