implement DatasetDict

[ArneBinder]

This PR implements a PIE DatasetDict similar to the HF version. However, it works as a container for pie Datasets and IterableDatasets and some methods are slightly different. List of implemented methods:

• __getitem__ *: returns Datasets or IterableDatasets
• from_hf_dataset: convert a HF DatasetDict or HF IterableDatasetDict to a PIE a DatasetDict (requires a document_type)
• to_json *: We just convert all documents with .asdict() and dump them with json.dump() to json line files
• from_json *: Loads the content dumped with to_json
• document_type (property): return the document type. Raises an error if no splits are available or the splits have different document types.
• dataset_type (property): return the dataset type, i.e. either Dataset or IterableDataset. Raises an error if no splits are available or the splits have different dataset types.
• map *: similar to the HF version. But the mapping callable (function) gets and needs to return a document. Also, checks function is an object that is derived from one of the following mixins and then calls the respective logic:
  • EnterDatasetMixin: calls function.enter_dataset(dataset_split, split_name) before processing the split
  • ExitDatasetMixin: calls function.exit_dataset(processed_dataset_split, split_name) after processing the split
  • EnterDatasetDictMixin: calls function.enter_dataset_dict(dataset_dict) before processing any split (end before handling any EnterDatasetMixin)
  • ExitDatasetDictMixin: calls function.exit_dataset_dict(processed_dataset_dict) after processing all splits (end after handling any ExitDatasetMixin)
• select: similar to HF version, but adds parameters start, stop, step that will be used to create indices, if available, and split to indicate which split should be modified
• rename_splits: rename the splits with a mapping dict
• add_test_split: cut a target_split out of a source_split by using train_test_split() from HF
• drop_splits: Drops splits from the dataset
• concat_splits: concatenate selected splits into a new split
• filter: filter a split via a function by using filter from HF. IMPORTANT: In contrast to map, the filter function gets the dict instead of a document as input because the PIE variant of Dataset.filter() is not yet implemented!
• move_to_new_split: similar to add_test_split, but the moved documents can be selected by a list of ids or via a filter function (uses PIE DatasetDict.select internally).
• cast_document_type: casts all dataset splits to a new_document_type

IMPORTANT: Methods marked with (*) differ from the HF syntax and semantic!

This PR also adds utils.hydra.resolve_target().