Generative Adversarial Network Art 🎨

Creating a Generative Adversarial Network to generate modern art and training it | Numpy, PIL, keras

Concept 🔎

Generative adversarial networks (GANs) are algorithmic architectures that use two neural networks, pitting one against the other in order to generate new, synthetic instances of data that can pass for real data.

Steps 🐌

1 - Get from the catalog of the Web Galery of Art (https://www.wga.hu/) some dataset

We set some features and get the URLs with the downloader.py script

form_feature = "painting"
type_feature = "landscape"

This parameters above gave us almost 3000 pictures.

Le golfe de Marseille vu de l’Estaque, par Paul Cézanne

2 - Resize the images from the dataset with resizer.py to 300x300

3 - Train the models with art_gan.py and see the results each 100 epochs in the output\ folder

Spell 💻

It is possible to use Spell (https://spell.ml/) to compute the program online

spell login
Spell upload wga.npy
Spell run python art_gan.py -t cpu -m uploads/art_gan_3000/wga.npy

Problems 🔧

Looks like the accuracies of both models tends to 100% very fast, and the results are mysterious.

Mean discriminator accuracy: 99.3514311650107, Mean generator accuracy: 99.41208541846206
10000 epoch, Discriminator accuracy: 100.0, Generator accuracy: 100.0

Ressources 📚

https://towardsdatascience.com/generating-modern-arts-using-generative-adversarial-network-gan-on-spell-39f67f83c7b4 https://machinelearningmastery.com/practical-guide-to-gan-failure-modes/ https://towardsdatascience.com/generating-abstract-art-using-gans-with-keras-153b7f11bd0 https://towardsdatascience.com/gan-ways-to-improve-gan-performance-acf37f9f59b