Yolov8 parse roboflow datasets

[IgorSusmelj]

Changes

• add explicit support for datasets that have wrong relative paths in the .yaml file such as

.
├── README.dataset.txt
├── README.roboflow.txt
├── data.yaml
├── test
│   ├── images
│   └── labels
├── train
│   ├── images
│   └── labels
└── valid
    ├── images
    └── labels

10 directories, 3 files

with data.yaml content of:

train: ../train/images
val: ../valid/images
test: ../test/images

nc: 3
names: ['0', '1', '2']

roboflow:
  workspace: ilyes-talbi-ptwsp
  project: futbol-players
  version: 9
  license: Public Domain

How it has been tested

• new unit tests

[MalteEbner]

Same fix is also used by ultralytics: https://github.com/ultralytics/ultralytics/blob/da9d8730f9f8a99fe6cd980fc927fd1005c15676/ultralytics/data/utils.py#L312

[MalteEbner]

While this fixes it for the case of roboflow datasets, it makes it fail for cases where the ../ is intended. Are you sure this is what we want?

Suggestion: Handle it the way ultralytics does it.

[MalteEbner]

LGTM

[MalteEbner]

Please keep the leading underscore. This makes it easy to find the corresponding test class for every class via STRG-F. The test class should always be named Test{ClassName} while keeping all leading and trailing underscores.
See our style guide: https://www.notion.so/getlightly/Python-testing-78e2dc49baf340178fa44ce20031c52e

Suggested change

	class TestYOLOv8BaseInput:
	class Test_YOLOv8BaseInput:

[MalteEbner]

Having a class within a class is something I haven't seen yet. I think it is ok. However, I would have preferred the keeping the usual naming of

test_get_categories_from_dict_format
test_get_categories_from_list_format
test_get_categories_from_yaml_block_format
test_get_categories_invalid_names

That would make the tests shorter and remove the extra class in a class. It would also allow to fin the tests corresponding to a method faster via STRG-F.

[MalteEbner]

Same for the other classes within classes

Feel free to resolve.