Graham Neubig

Thanks for the quick reply. `afe5d4734e179155852ea3a2f80c353f58c8b6ec` still encountered the same problem. Also, to make sure it wasn't the machine(s) that we were using, I also started an Amazon instance (p2.xlarge,...

Our preliminary tests indicate that this did indeed fix the nan errors. I'll re-open this if we encounter it during a full training run, but I think this is likely...

This is not supported yet, but it would be a nice feature to have.

Try `mkdir models` before running lamtram.

Hmm, has this bug resolved itself? It's a rather puzzling one, as it should just work.

Thanks for the report. I can't confirm this behavior. Can you give me more information about the version of KyTea you are using, the exact command you ran, the environment...

This is mainly an issue with upstream dynet, which has some tips on where to get started: https://github.com/clab/dynet/issues/92 I'd say that we're 90% of the way there, so a concerted...

Instead of "--operation gen" try "--operation nbest". I think this should output the confidence scores as well. (If this works, please close the issue)

So I think this is probably because BARTScore is taking too long to compute and timing out. * **Short term fix:** don't use bart_score, or use it only on smaller...

yes, I think sticking with other metrics is fine for now