[GSK-2355, GSK-2349, GSK-2373] DataLoader structure

[Hartorn]

No description provided.

[linear]

GSK-2355 DataLoader structure

[linear]

GSK-2349 Caching

Ideas:

• Caching locally the predictions for the whole dataset + transformation with different configs (facial_part) [to be discussed]
• Hashing and caching a la https://github.com/Giskard-AI/giskard/blob/56543d90147e4b7f934f5a2579e982524d1fcd69/giskard/models/base/model.py#L377-L392

[rabah-khalek]

Overall great work @Hartorn!

I had doubts about promoting the class-based (as opposed to a method-based) approach to define augmented-datasets (and soon sliced-datasets) but atm I can't see any downsides as the class-based one allows for more customisability and flexibility.

I mainly did some refactoring and conflict resolutions with main.
In particular:

• I removed facial_part as a parameter to model.predict() as this should be dictated by the passed dataset.
• I renamed the base classes into:
  • DataIteratorBase: The base class for all dataloader classes serving as skeleton
  • DataLoaderBase: The base class for all dataloader classes that implement loading methods
  • DataLoaderWrapper: The base class for all dataloader classes that serve as wrappers.

I organised the spin-off like DataLoader300W into loaders and CroppedDataLoader into wrappers.

To me it was the clearest way for now, let me know if you have other ideas.

TODO:

• Could you please remove the introduced ground_truths from PredictionResult? Even-though it might make things easier in some cases, but we should avoid mixing data-related attributes with model ones.
• clean up commented code + write docstrings
• I can take care of ensuring that all notebooks are updated before we merge.

TODO In a different PR:

• We should refactor the batching and sampling that @pierlj done in https://github.com/Giskard-AI/loreal-poc/pull/3/files -- but no need to do it in this branch, let's try merging this one asap.
• unit-tests

[rabah-khalek]

More or less minor comments

[linear]

GSK-2373 meta data handling in dataloader

• returning conform output from dataloader
• meta is a dict()

return res_img, res_marks, meta 

• res_marks, meta are optional