Training GAN for image inpainting

Now, it’s finally time to train a new GAN model for image inpainting. We can get the code from the original PyTorch implementationhttps://github.com/DAA233/generative-inpainting-pytorch. We will use the CelebA dataset as a training dataset for the experiment.

⚠️ The dataset is intended only for non-commercial research and educational use.

Model design for image inpainting

The GAN model for image inpainting consists of two generator networks (a coarse generator and a refinement generator) and two discriminator networks (a local discriminator and a global discriminator), as shown here:

Image $x$ represents the input image, $x_1$ and $x_2$ represent generated images by coarse and refinement generators, respectively. $x_r$ represents the original complete image and $m$ represents the mask for the missing part in the image.

The generator model uses a two-stage coarse-to-fine architecture. The coarse generator is a 17-layer encoder-decoder CNN, and dilated convolutions are used in the middle to expand the receptive fields. Assume that the size of the input image ($x$) is $3 \times 256 \times 256$, then the output ($x_1$) of the coarse generator is also $3 \times 256 \times 256$ ...