Time-Series-Generation

0. Overview

This repository provides a simple project on time series generation using GAN and PyTorch.

What this repository supports follows:

• Implementation of different models, LSTM-based and Conv-based models, for comparison
• Implementation of different loss functions, L2 and WGAN-GP, for comparsion
• Calculation of metrics proposed by TimeGAN (NeurIPS 2019) for quantitative evaluation

The models show bad performance. The performance depends upon the characteristic of dataset and selection of hyper-parameters heavily.

The dataset used here is Appliances Energy Prediction Data Set and can be found here.

1. Quantitative Evaluation

Model	Predictive Score↓	Discriminative Score↓
LSTM-based Generator with L2 Loss	0.0582	N/A
Conv-based Generator with L2 Loss	0.0466	N/A

The used metrics are simply adopted from the official repository of TimeGAN (NeurIPS 2019).

2. Qualitative Evaluation

LSTM-based Generator with L2 Loss	Conv-based Generator with L2 Loss

3. Run the Codes

1) Prepare Data

Please place the file to ./data/ and specify column that you want to generate.

2) Train

If you want to train Conv-based Generator with WGAN-GP loss,

python main.py --mode 'train' --model 'conv' --criterion 'wgangp'

To handle more arguments, you can refer to here.

3) Test

You can save plot and csv files then derive metrics on the test set. Please install tensorflow 1.15 to obtain corresponding metrics.

python main.py --mode 'test'

References

• Metrics for Quantitative Evaluation from TimeGAN (NeurIPS 2019)
• Data Pre-processing from WGAN_financial_time-series

Development Environment

- Ubuntu 18.04 LTS
- NVIDIA GFORCE RTX 3090
- CUDA 10.2
- torch 1.6.0
- torchvision 0.7.0
- tensorflow 1.15
- etc