Here are just a few records when I was first trying out a deep learning project. Including how to use packages to separate the parts of the project, what to do with each part of the project, what to do if using configuration files for management in a unified way, logging system and checkpoint management, and so on.
I have thought about and observed a lot of things thought this project, such as the relationship between batch size and learning rate, and so on.
You can see a better and more generalized version of the project that I completely thought about and realized in text-classification-trainer.