GANs-for-1D-Signal

Introduction

This repo contains pytorch implementations of several types of GANs, including DCGAN, WGAN and WGAN-GP, for 1-D signal. It was used to generate fake data of Raman spectra, which are typically used in Chemometrics as the fingerprints of materials.

If you use these codes, please kindly cite the this repository.

Requirements

• python 3.7.8
• pytorch 1.6.0
• numpy 1.19.2
• matplotlib 3.3.0

Experiment Result

DCGAN

WGAN

WGAN-GP

NOTE: RMSprop was used in the implementation of wgan-gp, rather than Adam, which was used in its original version, as it seemed like Adam didn't work well in my applications.

Comparison

Usage

data

You need to put all your data in the same folder in txt format, and make sure they are column vector.

network

The length of signal in my application is 1824, you need to modify the networks according to your data.

train

Run python file end with "train".

Reference

[1] Nathan Inkawhich. DCGAN Tutorial [EB/OL]. https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html, 2020-10-14.

[2] Yangyangji. GAN-Tutorial [DB/OL]. https://github.com/Yangyangii/GAN-Tutorial, 2020-10-15.

[3] mcclow12. wgan-gp-pytorch [DB/OL]. https://github.com/mcclow12/wgan-gp-pytorch, 2020-10-15.