Abstract

Image-to-image translation is used in a broad variety of machine vision and computer graphics applications. These involve mapping grey-scale images to RGB images, deblurring of images, style transfer, transfiguring objects, to name a couple. In addressing complex image-to-image translation issues, Generative Adversarial Networks (GANs) are at the forefront. StarGAN-v2 is a state of the art method for multi-modal multi-domain image-to-image translation that produces different images from a single input image over multiple domains. However, at a parameter count of more than 50M, StarGAN-v2 has a computation bottleneck and consumes more than 60G MACs (Multiply-Accumulate Operations to calculatecomputation expense, 1 MAC = 2 FLOPs) to create one 256×256 image, preventing its widespread adoption. This thesis focuses on the task of compressing StarGAN-v2 using knowledge distillation. Using depthwise separable convolutional layers and reduced channels for intermediate layers, we develop efficient architectures for different StarGAN-v2 modules. In a GAN mini-max optimization environment, the efficient networks are trained with a combination of different distillation losses along with the original objective of StarGAN-v2. Without losing image quality, we reduce the size of the original framework by more than 20× and the computation requirement by more than 5×. The feasibility of the proposed approach is demonstrated by experiments on CelebA-HQ and AFHQ datasets.