I have been playing with Pytorch recently and found out that Pytorch is a much easier tool for me to start a machine learning project than Tensorflow. Pytorch uses a different design with Tensorflow and does not need to setup sessions or placeholders, so I feel the setup of Pytorch models is more intuitive. Pytorch’s graph is called Dynamic Computation Graph, which is generated on the fly and allows users to manipulate the graph on the go.
In order to make myself familiar with Pytorch, I started a small project about foot image classification. The basic idea is to train a classifier to recognize whether a pressure image belongs to a left or a right foot. Some of our research works incorporate with floor sensors, but floor sensor itself would not be able to tell if a blob is part of the left or right foot. Recognizing left/right foot image is very helpful when locating the separate feet from the foot placement and when we wanted to analyze separate feet such as high/low pressure sections, arch index, etc.