Preprocess the RVL-CDIP dataset into a grouped by category format

The RVL-CDIP should be downloaded from the original website - https://www.cs.cmu.edu/~aharley/rvl-cdip/. This repository is particularly useful if you want to train a model on all images without worrying about benchmarks (it does not keep the original train/test/val split!). For example to train a network for later use for transfer learning :).

0. Move the downloaded file in a folder which has a lot of disk space.

1. run tar -xvzf "./rvl-cdip.tar.gz"

2. The directory should look something like the image, without the dataset folder (that one is created automatically by this script)

3. Use the compose.py to build the per category dataset. Note it might take a while.

python compose.py

4. Done