Dataset.from_pandas with a DataFrame of PIL.Images

[lhoestq]

Currently type inference doesn't know what to do with a Pandas Series of PIL.Image objects, though it would be nice to get a Dataset with the Image type this way

[mariosasko]

A duplicate of #4796.

We could get this for free by implementing the Image feature as an extension type, as shown in this Colab (example with UUIDs).

[ZachNagengast]

+1 to this
Calling this line with a df that contains a PIL image (as they are returned from load_dataset)
ds = Dataset.from_pandas(df)
Results in this error:
ArrowInvalid: ('Could not convert <PIL.PngImagePlugin.PngImageFile image mode=RGB size=1024x1024 at 0x2B41F2D70> with type PngImageFile: did not recognize Python value type when inferring an Arrow data type', 'Conversion failed for column image with type object')

[juanxo90]

I found something that can be used as solution.

I have the same problem when I've try to load the images from a pamdas dataset

If you have all on a pandas dataset try
Dataset.from_dict( your_df.reset_index(drop=True).to_dict(orient='list'), split=set_your_split)

And this avoid the error