Paul O'Leary McCann

There may still be a few rough edges in the code, but I believe at this point it's ready for review, and strange types / lack of clarity should be...

You can also test a fully working component with the project here. https://github.com/explosion/projects/pull/101

Tests are failing due to an error in thinc related to trying to unpack a dlpack twice. The failures are in the `xp2torch` function and affect components besides coref. I...

Thanks to @adrianeboyd for pointing out this error is due to having an old Torch and new numpy, which aren't compatible. I have upgraded the Torch version used in the...

That seems to have fixed the tests.

> I assume you're waiting on the spacy docs to decide how to add this to the README? Yes, it wasn't clear if details should go here or in the...

If the note in the docs is OK, I think that all the concerns about this have been addressed, and this should be ready to be merged.

Yes, I think it makes sense to follow up on performance later. To summarize the results of profiling, it looks like most of the time is in Torch layers, and...

OK, now that performance profiling is done for the time being I think this is ready to merge?

OK, there was a small issue with the optimizations on Windows, but now the tests have cleared (including project tests on Windows) and I was able to train a whole...