Computer Vision News - January 2022

14 Deep Learning Tool DRAWING NEURAL NETWORKS by Marica Muffoletto (Twitter ) Hi all! I am delighted to start yet another year with the RSIP Vision community. I wish both regular and new readers 12 months filled with happiness, joy and “more importantly” exciting computer vision content! We begin 2022 with an article proposed by one of our own readers, discussing a hot and useful topic in the AI and Deep Learning community: how to draw neural networks. As we all know, this field is growing at light speed and diagrams of Neural Network architectures are commonly found in most computer vision papers and proudly shown everywhere in conferences. Hence, it becomes essential to add this to our skillset: it could be used to show a schematic of your algorithm to the team, to be included in a paper, in a poster or a talk. Images of Neural Networks are sometimes even more explanatory than the corresponding code or description. They are usually one of the best methods to represent a model. A fast, easy solution that is widely employed is PowerPoint: the Microsoft suite offers shapes, charts and icons that can be organised to mimic the common diagram, but today we are going to focus on more complex tools. Let’s start from picking a pre-defined model as an example that we will keep throughout the article and aim to draw using different tools. I chose VGG16, an architecture for classification and detection which won the ILSVR (ImageNet) competition in 2014. Most famous deep learning libraries PyTorch & Keras have common architectures stored as classes and these can be called and printed with a summary of the layers and the output shapes. The table below shows how to achieve this and what output is displayed for both cases.