The Dataloader collates (with this?) the samples such that the labels are also collated within each key of the sample dictionary. I mistakenly thought I could extract a specific sample meta by simply using:
ind = 12
meta_of_one_sample = {key : labels[key][ind] for key in labels}
it gets more complicated when the sample dictionary has a value that is a nested list, then the nested lists collates too and the code above returns nonsense.
Is there an uncollate function?