Computer Vision News - January 2020

2 Summary We Tried for You 4 Every once in a while, there comes a library that changes the world of Deep Learners. At the beginning this was TensorFlow, then Keras , and now it is PyTorch . PyTorch was created to give the deep learning community an easy and pythonic way to build and design their neural network architectures. It gives the user an easy to use framework by Amnon Geifman In the task of neural style transfer, we would like to take two images: a style reference image (the abstract image below) and a content image (the shoes image below); and blend them so the content of the new image looks like the content image and the style of the new image looks like the reference image . This is done by optimizing the output image, Getting started Features of PyTorch like Keras, with the flexibility allowed by TensorFlow, which makes it a very powerful framework. In this article we introduce several useful features of PyTorch that demonstrate the efficiency of this library . We will show these features through an implementation of neural style transfer . "At the beginning this was TensorFlow, then Keras, and now it is PyTorch" in such a way that it will contain the content statistics of the content image and the style statistics of the style reference image. These statistics are extracted from the images using a pretrained convolutional network. For our purpose, it helps to demonstrate how easy it is to get inside the network and perform manipulations with PyTorch. So, let’s begin. In this article we introduce several useful features of PyTorch that demonstrate the efficiency of this library