Educational Resources for Remote Comparative Study of Generative Adversarial Networks

Studying machine learning methods for working with images is important in the education of future information technology specialists. In this regard, software tools are considered for interactive educational study and comparison of the methods for training generative adversarial networks for image inpainting, image super-resolution, and global illumination rendering for 3D scenes. The educational resource is represented as a configuration of TensorFlow and Jupyter Notebook environment tools. It enables remote study of the impact on image inpainting quality of the addition of residual blocks to the generator network, adjustments to the networks’ learning rate, and the proportion of damaged pixels in the input data. Additionally, the educational resource supports interactive tools for remote studying of the super-resolution method in comparison to bicubic interpolation. It also explores the effect of using various upsampling methods and applying different components of the loss function. Furthermore, the method of global illumination rendering from rasterized maps is studied in comparison to the ray tracing method. It allows for the study of the effect of adding a low-resolution ray-traced image to the input using the proposed educational resource. The workbench is utilized in the laboratory practice for the "Methods of Machine Learning" course for students of Applied Mathematics and Computer Science.