加入CODE CHINA

· 不限速    · 不限空间    · 不限人数    · 私仓免费

免费加入
readme.md

InfoGAN for temporal series generation

Introduction

Use infoGAN to generate multi-dimentional temporal sequences.

In MNIST, continuous noise can represent character width and rotation angle. But in here, their meanings are not quite intuitive.

In MNIST, modifying continuous noise would not change the classtype of generated samples. But in here, it happens.

See codes for MNIST at https://github.com/SongDark/GAN_collections.

Datasets

Name Link Class Dimension Train Size Test Size Truncated
CharacterTrajectories Download 20 3 1422 1436 182

Generate npz file

Unzip CharacterTrajectories.zip at data/CharacterTrajectories, then run dataprocess.py.

    python dataprocess.py

Generated Images

Random noise, Random discrete code, Fixed continuous code

Epoch 0 Epoch 200 Epoch 500

Specified condition, Random noise

Epoch 0 Epoch 200 Epoch 500

Fixed noise

It seems that the generated sequences are not corresponding to their one-hot labels.

label=14(s) label=16(v) label=16(y)

Reference

https://github.com/hwalsuklee/tensorflow-generative-model-collections/blob/master/infoGAN.py

https://github.com/buriburisuri/timeseries_gan/blob/master/train.py

项目简介

🚀 Github 镜像仓库 🚀

源项目地址

https://github.com/songdark/timeseries_infogan

发行版本

当前项目没有发行版本

贡献者 1

S SongDark @SongDark

开发语言

  • Python 100.0 %