Use the same train-val-test split

[hermancollin]

In the preprocessing file, the data is first converted to YOLO format, then COCO. The problem is that the data is shuffled at both steps, resulting in random non-matching data splits. Both splits should be identical, otherwise how could we compare both methods???

There should be a JSON or YML file with a hardcoded data split. Both preprocess_data_coco() and preprocess_data_yolo() should split the data identically.

[hermancollin]

We should hardcode the data split in a JSON file inside this repo. It should look like this:

{
   train: [
      "sub-rat1_sample-[...].png",
      "sub-rat2_sample-[...].png",
      [...],
   ],
   val: [
      [...]
   ],
   test: [
      [...]
   ]
}