compatibility with spaCy 3.0 ?

[d5555]

Is there going to be neuralcoref compatible with spaCy 3.0 ?

[svlandeg]

Yes! This is definitely on our road map :-) I'll actually pin this issue because I'm sure others are interested as well.

[asterbini]

Plus 1

[asterbini]

After some rough paches I am able to start neuralcoref with spacy 3 but I obviously get the error
''Trying to read a Model that was created with an incompatible version of Thinc''
Is there a way to convert the models to the new Thinc version used in Spacy 3?

[svlandeg]

@asterbini : Due to the many internal changes between Thinc 7 and 8, and spaCy 2 and 3, I'm not sure this is going to be feasible. See also this discussion at the spaCy forum.

We have now started working on a coref component native in spaCy v3 though. This won't be finished tomorrow, but we're making progress. We're going to reuse what we can from this library (with attribution, ofcourse) and make sure it trains and runs with native Thinc models :-)

[asterbini]

I was starting to look for a way to export the neuralcoref model in spacy2 as text, and try to load it in spacy3 :)
From your words this seems to be useless, thanks for letting me know.

[svlandeg]

I'm not a 100% sure though. If you have a working PyTorch model, that might operate with Thinc's new shim layers, but you'll have to write a custom component to wrap it in spaCy...

[asterbini]

But why the neuralcoref models should depend on the spacy3 models implementation? From a cursory glance to the NeuralCoref annotator class it seems to me that it reads only properties of the doc tokens that have not changed from v2 to v3. Then unless thinc 8 is changed radically wrt 7, (and by mapping Relu to ReLu and Linear to Affine) perhaps there is a way to convert the binary data model to thinc 8.
(and yet, as I am a newbie of thinc and neuralcoref an I am sure I could have missed something)

[KTRosenberg]

@asterbini : Due to the many internal changes between Thinc 7 and 8, and spaCy 2 and 3, I'm not sure this is going to be feasible. See also this discussion at the spaCy forum.

We have now started working on a coref component native in spaCy v3 though. This won't be finished tomorrow, but we're making progress. We're going to reuse what we can from this library (with attribution, ofcourse) and make sure it trains and runs with native Thinc models :-)

Out of curiosity since I'm trying to integrate Spacy and Neuralcoref into a time-sensitive project, is the move to Spacy 3 going to take on the order of a couple of weeks, a couple of months, or a year? I'm not familiar with the amount of work necessary to do something like this.