Multi-OctConv: Reducing Memory Requirements in Image Generative Adversarial Networks

Generative Adversarial Networks (GANs) for image generation of human faces have provided excellent results in recent years. However, we were able to identify a common problem among them: high memory usage in their training phase due to the convolutional encoder architecture used in these models. We address this issue by replacing the traditional convolutional layers in a model by what we call a Multi-Octave Convolution (M-OctConv) without modifying its architecture. An advantage of this method is that it can be easily combined with traditional memory reduction techniques, such as pruning. We evaluate our proposition on StarGAN model achieving up to 40% of memory usage reduction without affecting the quality of the generated images.