Is your feature request related to a problem? Please describe.
During evaluation, now we use prepare_batch to extract image and label fields if existing the key. But maybe we don't want to load label for inference and didn't apply transforms, so need to ignore it for this case.