I did use paperspace with duckduckgo image downloader. I tried to model one of my idea in smaller scale (3 categories out of 30-40 categories).
However what I did these 3 categories look like each other so my error rate is 0.15 which based on Jeremy Lecture is too high. However I extract the model and I can make a web app with Django and Javascript in 1-2 days.
What can I do to decrease these error rate ?
My first plan is to add all of the categories
My question is how I can get better images ? Is it possible to change to see which categories of images are in 22000 (big image data set) and so If they have the categories that I want to use I can use their image ?
I also want to thanks Jeremy and everyone in fastAI community that are supper helpful.
It felt so good that I made my idea however it is the beginning for me.