Hey guys, I just finished my project on Image classification using the Covid-19 dataset. It was great fun! I used a somewhat different approach though than given in the project. I wanted to avoid using the same data I trained on to do my predictions later an set aside a completely unseen set of data. So I reordered the images in one larger folder with all ‘Covid’ ‘Normal’ and ‘Pneumonia’ images an then sampled using the splitfolders package I found on pypi:
The complete code and data can be found under:
That way I was able to get up to around 50% precision and recall while converging with over 0.96 on the validation AUC. I’m not completely happy though. According to the generated graph my accuracy and AUC sometimes do huge jumps in values. Is there anything I can do to improve stability here?
I’d absolutely love to get some input from the community here so feel absolutely free to critizise!
Excited to hear from you guys.