Hi,

There's existing discussion around a more extendable RNN API, which as of yet hasn't moved forward. Given that, and that PackedSequences exist, would a reverse_padded_sequence function be useful, or do you just see it as being unnecessary clutter?

In particular it'd take in a Variable of padded sequences along with sequence lengths (and a batch_first option) and produce a Variable of padded sequences that are reversed according to their lengths, mimicking reverse_sequence in TensorFlow.

My thought is that this option would be a useful alternative to PackedSequences in the case of bidirectional RNNs. (PackedSequences seem to be primarily motivated by cuDNN, whereas in a research setting they might introduce more bookkeeping than they're worth.)

cc @ezyang @gchanan @zou3519