TimeGAN-PyTorch

Unofficial implementation of TimeGAN (Yoon et al., NIPS 2019) in PyTorch 2.

Jinsung Yoon, Daniel Jarrett, Mihaela van der Schaar, "Time-series Generative Adversarial Networks," Neural Information Processing Systems (NeurIPS), 2019

Original Codebase: https://github.com/jsyoon0823/TimeGAN.git

Data Set Reference

• Stock data: https://finance.yahoo.com/quote/GOOG/history?p=GOOG
• Energy data: http://archive.ics.uci.edu/ml/datasets/Appliances+energy+prediction

Version Notes

The model was implemented and tested using Python==3.11.9. Further, the following modules were utilized (see Requirements File):

matplotlib==3.10.3
numpy==2.3.1
pandas==2.3.0
scikit-learn==1.7.0
torch==2.7.1
tqdm==4.67.1

Usage

To conduct the experiments, the easiest way to get started is by cloning this repository and use the notebook.

Alternatively, run it from the terminal.

py main_timegan.py --data=data/stock_data.csv --seq_len=24 --module=gru --hidden_dim=24 --num_layers=3 --epochs=10000 --batch_size=128 --metric_iteration=10 --learning_rate=1e-3

Results

Stock Data

Results obtained from the notebook.

1. Discriminative Score

#Compute discriminative score
discriminative_score_metrics(data_train, data_gen, device)

100%|██████████| 2000/2000 [03:17<00:00, 10.10it/s]

0.23260572987721695

2. Predictive Score

#Compute predictive score
predictive_score_metrics(data_train, data_gen, device)

100%|██████████| 5000/5000 [07:20<00:00, 11.35it/s]

0.04312044765531347

3. Visualization