Raise a clear error for values that cannot be encoded as an image

[LeSingh1]

Putting an unsupported value in an Image() column fails with an error from deep inside the encoder rather than a message about the image:

from datasets import Dataset, Features, Image

Dataset.from_dict({"im": [123]}, features=Features({"im": Image()}))

AttributeError: 'int' object has no attribute 'get'

Same for a float, a bool, a set, or any other unsupported object. A 0-dimensional numpy array fails differently but just as opaquely:

Image().encode_example(np.array(5))
IndexError: tuple index out of range

Cause

Image.encode_example handles str, Path, bytes, np.ndarray and PIL.Image.Image, and then falls straight through to the mapping branches:

elif value.get("path") is not None and os.path.isfile(value["path"]):

so any value that is not one of the supported types and not a mapping raises AttributeError on .get.

The Video feature already guards exactly this case:

# src/datasets/features/video.py
else:
    raise TypeError(f"Unsupported encode_example type: {type(value)}")

Image simply never got the same guard.

Fix

• reject values that are neither a supported type nor a mapping with the same TypeError: Unsupported encode_example type: ... message Video already uses, so the two features report this identically;
• reject 0-dimensional arrays in encode_np_array with an explicit ValueError instead of letting the shape indexing blow up.

Mapping is used rather than dict so that any mapping type keeps working. Mappings that are missing both path and bytes still raise the existing, already explicit ValueError, and every currently supported input is unchanged — str, Path, bytes, dicts, PIL images and 1D/2D/3D arrays all still encode exactly as before.

Tests

• test_image_feature_encode_example_unsupported_type covers int, float, bool, set and a plain object;
• test_image_feature_encode_example_zero_dimensional_array covers the 0-d array through both Image.encode_example and encode_np_array;
• test_image_feature_encode_example_missing_keys pins the existing ValueError for a mapping without usable keys, so this change cannot swallow it.

The first two fail on main (6 parametrizations); the third passes before and after and is there as a regression guard.

tests/features/, tests/test_formatting.py, tests/test_table.py and tests/test_arrow_dataset.py: 1068 passed, 0 failures.