Matt

@lhoestq Are you okay with me merging this for now?

Right now, the `auto_rename_labels` behaviour happens in all cases! Making it an option is the first step in the process of disabling it (and moving the functionality to `transformers`) and...

This has so much potential to be great! Also I think you tagged some poor random dude on the internet whose name is also Joao, lol, edited that for you!

cc @sayakpaul here too, since he was interested in our new approaches to converting datasets!

Yeah, it looks like in its current state the tfio support for `Feather` is incomplete, so we'd end up having to write a lot of it, or do a conversion...

@lhoestq Although I think this is a dead-end for now unfortunately, because of the limitations at TF's end, we could still explore automatic conversion to TFRecord, or I could dive...

> If to_tf_dataset can be unbatched, then it should be fairly easy for users to convert the TF dataset to TFRecords right ? Sort of! A `tf.data.Dataset` is more like...

Investigating now, pinging @nateraw and @gante on this one too! I've confirmed that there is an underlying issue here where `model.save` is failing on Keras sequential/functional models containing HF models...

Actually, wait, I think the issue may be with some of the example code! Replacing `hf_result = transformer_model([input_word_ids, input_mask])` with `hf_result = transformer_model(input_word_ids, input_mask)` causes one of the examples to...

Yeah, this is something we should either support fully, or be very clear that we don't support - something that works half the time but breaks when you try to...