Repository for playing/studing with Generative Adversarial Network.
In 2023, I made my first attempt at DCGAN training.
train_font_dcgan.py trains Deep convolutional GAN with JIS Level-1 Kanji(2965 characters) and 8 different font files.
# download font files
python -m font_resource.download_google_fonts
# training
python -m playground.gan.train_font_dcgan --data-dir ./tmp/dcgan --model-dir ./tmp/dcgan
The generated images are saved in tmp/dcgan.
Here is a generated image.
There are several Kanji characters that can be read such as 泊, 枉, 恫, 花, etc. This may be overfitting the training data.
The architecture is a bit more complex than the reference implementation. Two discriminators are used: low level feature discriminator and high level feature discriminator.