An example of how to use fastai to classify newspaper pages.
The combination worked considerably better for various classification tasks on the Berliner Börsen-Zeitung than using a default resnet18:
- xresnet18
- mixed precision training
- lessw2020's implementation of Ranger
- downscaling to 448 pixels