Closed robertostling closed 7 years ago
There was a bunch of conflicts with Jörg's branch. I have hopefully resolved these now, but the code needs a bit more testing before I merge into master. Unfortunately the next TextEncoder breaks compatibility with old model files, so I will probably make a new branch to support those models.
This removes some earlier functionality, like the batched beam search (replaced by a per-sentence beam search with coverage penalty).
The neural computation in the new beam search is still batched. The difference in the implementations is that the sorting and pruning of hypotheses (previously done with numpy) is now done in a different way (with native python).
Great! I have merged this with the joerg branch. Since there were a lot of conflicts I leave the code in coverage_penalty for testing for a while, then we'll hopefully make this the master branch.
This removes some earlier functionality, like the batched beam search (replaced by a per-sentence beam search with coverage penalty). Some components should possibly be merged into BNAS.