Sockeye vs gluonNLP


I’m looking for the easiest MXNet-backed solution to train a Transformer seq2seq model on custom seq-2-seq pairs. Shall I use gluonNLP or Sockeye? Happy to know key differentiators of each toolkits

Given that Sockeye is specifically designed for Seq2Seq I’d definitely look here first.

As a disclaimer I haven’t actually used the project before, but the documentation looks great and I like that data preparation and inference are covered in detail. I can also see that Sockeye has a number of interesting contrib packages, such as autopilot and vistools (for the visualising beam search).

GluonNLP would be preferred if you were thinking of doing lots of adjustments to the models and wanted more fine grained control over training. Since these things are typically easier to do in Gluon.