This is my project to learning image recognition with using of neural network. Whole code is upload in this repo. I split it to two Jupyter Notebooks. One is for preprocesing, and the second is for learning. If you want something more, pleas write to me! The dataset i use is HAM10000. Full information and downloadable files you can get from web page linked bellow.
Basics information:
Learning neural network on different numbers of pictures in each class(like class 1 =8000 images and other 6 classes ~100 images each) is bad because your neural network learn to predict only one class. It leads to wrong evaluate metrics. In other way if you create a lot new pictures from "Data generator" you increase learning time. Maybe if you create more epoch you can get more accurate model. But as i see the validation accuracy decreas by second epoch.
Summary of summary, now:
"I know that I know nothing" - Socrates