README.md

keragan

Keras implementation of GANs

This library provides some simple infrastructure to define and train Generative Adversarial Networks in Keras. It can also be used from the command line.

Installation

The simplest way to install is:

pip install keragan

Some Images produced by keragan

Images were trained on WikiArt.org dataset taken from here.

Minimalistic Landcape, 2020	Spring Dawn, 2020	Water Still 2, 2020
---	---	---

Training GAN from the Command-Line

To start training, you can use the following command-line:

python -m keragan.trainer c:\dataset --size 1024 --model_path .\models --samples_path .\samples --latent_dim 100 --epochs 1000

You can find out more about other parameters by calling the program with --help option.

Important things to note:

• --size must be power of 2, suitable values are 64, 128, 256, 512 and 1024. Higher resoltions are likely not to give good results.
• You can use --lr to set learning rate, default value is 0.001. Smaller learning rates yield better results, but may significantly increase training time.

Generating Images

Once you have trained the model, you can use the generator model to produce new random images. To do that from a command line, you can use the following:

python -m keragan.generate --file ./models/gen_1100.hdf5 --out ./samples --n 100

Use --help option to find out more about different options.

Architecture

The library is structured around few core classes:

• GAN is used to represent a GAN, with generator and discriminator fields that define corresponding networks. GAN itself is abstract, and any subclass should define create_generator() and create_discriminator() functions. This class is also responsible for loading/saving networks to disk, and it can also generate sample images using sample_images method.
• DCGAN is currently the only subclass, implementing Deep Convolutional GAN.
• ImageDataset is a class defining the process of loading initial dataset from disk, resizing it to specified size, filtering out bad images, etc.
• GANTrainer is responsible to training a GAN, i.e. running epoch loop and periodically storing samples and network weights to disk.

The actual training code looks like this:

    gan = keragan.DCGAN.from_args(args)
    imsrc = keragan.ImageDataset.from_args(args)
    imsrc.load()
    if args.sample_images:
        imsrc.sample_images()

    train = keragan.GANTrainer(image_dataset=imsrc,gan=gan,args=args)
    
    def callbk(tr):
        if args.visual_inspection_interval and tr.gan.epoch % args.visual_inspection_interval == 0:
            res = tr.gan.sample_images(n=2)
            fig,ax = plt.subplots(1,len(res))
            for i,v in enumerate(res):
                ax[i].imshow(v[0])
            plt.show()

    train.train(callbk)