Deep Convolutional GAN trained to generate landscape paintings. (work in progress)
I am still trying different training configurations (training GANs is not easy!), but this is a sample of generated images by the best model so far.
The architecture of the network is based on the PyTorch implementation from the paper 'Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks' by Soumith Chintala. https://arxiv.org/abs/1511.06434
Some important changes are:
- Doubled image size: now 128x128 instead of 64x64 (adding a layer in both networks)
- Unbalanced G/D channels: now ngf=160, ndf=40 instead of ngf=64, ndf=64 (to give an advantage to the generator)
The model was trained on a Jupyter Notebook, using Kaggle Kernels.
The dataset used for training was scraped from https://www.wikiart.org/ with a python scraper built specifically for the task. It includes thousands of paintings with landscapes as main theme.
Both the training notebook and the scraper are available in this repository.