Modular GANs - Test bench for my GAN related Experiments

This repo is a testbench for my immature GAN related experiments. I plan to implement and carefully understand various GAN related research papers.

Generated images after 1500 epochs:

About

• Dataset: Anime girl faces (weeb)
• train.py for training the model. Replace load_path = "logs/models/1250.torch" with the latest model to continue training. TODO: Take this from command line using argparse.
• config.ini contains all the hyperparameters and other training details.
• networks.py contain the Generator and Discriminator network classes.
• load_data.py is the dataloader.
• 1500 epoch took ~16hours on my GTX 1060 6GB, i7 6700 and 16GB RAM.

Todo (in order of priority):

• Basic GAN
• Historical weight averaging (Tim et. al.)
• TTUR (works surprisingly good) Heusel et. al.
• Use config files.
• Latent walk.
• Train Time comparison.
• Load and save model for continuous training.
• Current GAN for general size and not just 64x64 pixels.
• Implement the FID metric.
• Other GAN losses like WGAN, RaLSGAN.
• New dataset and C-GANs.
• Progressively growing GANs - ProGAN, BigGAN, StyleGAN.
• More TODOs...
• Add comments.

Latent walk examples:

End notes

• Thank you kaggle for organizing the Generative Dog Images competition because of which I learned everything about GANs.

• http://www.thiswaifudoesnotexist.net/ wish I made this before them.