How would a decoder be implemented?

The original decoder from Kiros is here: https://github.com/ryankiros/skip-thoughts/tree/master/decoding

How would this be implemented in this repo?

Start by reading the original decoder under the linked skip-thoughts `decoding` path and comparing it with the code already present in this repository's PyTorch/Torch7 port. Identify the local model entry points that correspond to the original implementation, then define the decoder's expected inputs, outputs, and validation before choosing an implementation approach. Done means the decoder is integrated into this repository and its behavior matches the original model.

Written by the indexing model from the issue text.