In this series, Artem Nazarenko, Computer Vision Engineer at Everypixel shows you how you can implement the architecture of a neural network. In the first part, we were talking about the working principles of GAN and methods of collecting datasets for training. This part is about preparing for GAN training.
Loss Functions and Metrics
Attention. At this point, we deviate from the reference article. We take the loss function to solve the segmentation problem. Generation of attention maps (masks) can be considered as a classic image segmentation problem. We take Dice Loss as the loss function. It is well resilient to unbalanced data.