Fairseq: What is incremental_state?

[robertodessi]

I am trying to understand the decoder in the fconv architecture and I am wondering what does the incremental_state variable do and what values does it hold when doing inference? Is it not used at all at training time?

Thanks in advance

[myleott]

Please see my comment here: https://github.com/pytorch/fairseq/issues/166#issuecomment-392259850

You are correct that it is not used during training time, only inference time.

[ankur6ue]

I wrote up a blog post that describes some of what the incremental_decoder does:
http://www.telesens.co/2019/04/21/understanding-incremental-decoding-in-fairseq