Simplest way to use the wikitext103/wt103 model/weights/state

I would like to try out some semantic textual similarity with the wt103 model. All I want to do is pass some tokenized text in and get out a vectors I can do e.g. cos similarity on. I’d like to do this without specifying training/val paths since there are none, and without any additional training, which is different from all the tutorial situations. Arvind-cp great quora notebook is helpful and very relevant, but generally it copies the tutorials.

I ask for a few reasons: I suspect this minimalist model loading can be done in a few lines of code, and I haven’t found them yet. And, this is how people might want to set things up in production / a server.