Codebase for "Time-series Generative Adversarial Networks (TimeGAN)"

Authors: Jinsung Yoon, Daniel Jarrett, Mihaela van der Schaar

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

Paper Link: https://papers.nips.cc/paper/8789-time-series-generative-adversarial-networks

Contact: [email protected]

This directory contains implementations of TimeGAN framework for synthetic time-series data generation using one synthetic dataset and two real-world datasets.

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

To run the pipeline for training and evaluation on TimeGAN framwork, simply run python3 -m maintimegan.py or see jupyter-notebook tutorial of TimeGAN in tutorialtimegan.ipynb.

Note that any model architecture can be used as the generator and discriminator model such as RNNs or Transformers.

Code explanation

(1) data_loading.py - Transform raw time-series data to preprocessed time-series data (Googld data) - Generate Sine data

(2) Metrics directory (a) visualizationmetrics.py - PCA and t-SNE analysis between Original data and Synthetic data (b) discriminativemetrics.py - Use Post-hoc RNN to classify Original data and Synthetic data (c) predictive_metrics.py - Use Post-hoc RNN to predict one-step ahead (last feature)

(3) timegan.py - Use original time-series data as training set to generater synthetic time-series data

(4) main_timegan.py - Report discriminative and predictive scores for the dataset and t-SNE and PCA analysis

(5) utils.py - Some utility functions for metrics and timeGAN.

Command inputs:

• data_name: sine, stock, or energy
• seq_len: sequence length
• module: gru, lstm, or lstmLN
• hidden_dim: hidden dimensions
• num_layers: number of layers
• iterations: number of training iterations
• batch_size: the number of samples in each batch
• metric_iterations: number of iterations for metric computation

Note that network parameters should be optimized for different datasets.

Example command

$ python3 main_timegan.py --data_name stock --seq_len 24 --module gru
--hidden_dim 24 --num_layer 3 --iteration 50000 --batch_size 128 
--metric_iteration 10

Outputs

• ori_data: original data
• generated_data: generated synthetic data
• metric_results: discriminative and predictive scores
• visualization: PCA and tSNE analysis