📚 Documentation

There is a min_sample_size argument in audio pretraining task with confusing description (min sample size to crop to for batching). It is used only to set internal min_length property of dataset to filter small examples.

I propose to remove min_length property in favor of min_sample_size in audio datasets and improve description of this parameter.

I prepared branch, which I can submit as PR if this issue makes sense: https://github.com/pytorch/fairseq/compare/master...gazay:min_max_sample_size?expand=1