I post certain scripts I wrote for a fellowship.ai 3-months project. It was only a computational success, but not modelling success. I only knew very few python commands when I started the project. I did not even know what channel is. Eventually I built the skills of reading the codes quickly.
The purpose of the research is to apply Articifical Intelligence (AI) approach to classify X-ray images of Spinal Implants from eight manufacturers. The coding of outcome/classification was not certain, as some images were verified by radiologist (sometimes with annotations) and some from social media. I eliminated those seemingly gathered from social media. It left me with 452 images demonstrating diverse qualities and imbalance classes. One can expect that I got poor results, although I tried literally almost all possible techniques, including transfer learning. I did not try gridsearch techniques. I doubt that gridsearch is a magic bullet. We need to have a reasonable good model in order to obtain a desirable result. The best I could get was about 29% ish by implementing a VGG-19 and 56% ish by building a Siamese Network. Various authors claimed having successes with VGG-based models. In most occasions, the models did not learn. I got similar outcomes with models with various filter sizes, but still could not emerge a reasonable model.
I bet the scripts more or less cover the full-stack service - editing, cleaning.., and modelling. They may be of use to others. I did not apply subclass to build a model. It is easier to avoid mistakes by using Sequential or Functional APIs.
I may mention that I acquired most of the skills from Dr. Adrian Rosebrock's site and also Dr. Jason Brownlee's site.