dekeijzer / sron-dcgan

Using deep convolutional generative adversarial networks for the inpainting of exoplanetary atmosphere spectra

This repository is highly in development. A proper README will follow in the future.

SRON-DCGAN

Using deep convolutional generative adversarial networks (DCGANs) for the semantic inpainting of exoplanetary atmosphere spectra

Introduction

The ideas and models within this repository are bassed off of ExoGAN, link. ...

DCGAN

The DCGAN architecture from 'Unsupervised Representational Learning with Deep Convolutional Generative Adversarial Networks (2016)' is used and modified to work with 1x32x32 (Channel, Height, Width) shaped images.

Atmospheric Spectra and Parameters Array (ASPA)

Each image contains a Atmospheric Spectra and Parameters Array (ASPA).
The data used to create the ASPAs (currently) originates from TauREx and is released within the ExoGAN repository.
An example spectrum would be as follows.

The spectrum along with the parameters like planet mass, planet radius, planet temperature and gas mixture ratios are encoded into an ASPA.
These ASPAs are then used to train the DCGAN.
After ~8 hours of training on a Tesla M10 using 0.25 % of the original ExoGAN training set, the following result is gathered.

Semantic inpainting

GAN trained for 55 epochs (batch sized 64), on 0.1*25% of the first 50 ExoGAN chunks.
Inpainting took 5000 iterations, loss is 9.5.
Note that this inpainting is on an ASPA from the training set.

Results

...