API stability, plus docs and examples?

I’ve been hemming and hawing about adding pre-trained/trainable NN models into textacy, partly because I’ve been waiting to see how thinc develops rather than committing early to, say, keras. Do you have any sense of when thinc’s NN API will be stable? And do you plan to add more documentation and examples? Your latest blog post is great, but the code seems to very much be a work-in-progress.

As always, thanks for all your great work. I look forward to seeing how this package matures!

[honnibal]

If you’re trying to get stuff done I think you’ll be best off using PyTorch for now. If you’d be keen to hack on Thinc as well as your usage code, I would give a very different answer 😃.

Working in PyTorch is definitely better than waiting, though. The core NN code is short in all of these frameworks, so if you like Thinc later you can always port your stuff.

Personally I already feel much more productive in Thinc than I did in Keras, but there’s a huge “it’s my own code” bonus there. The priority with Thinc is very much to scratch our own itches and be a good dependency for spaCy. I expect the result will be useful to people beyond spaCy, but it’s not what’s on our mind right now.

[bdewilde]

@honnibal I’ve been watching progress on the develop branch… Very excited! Thanks for all your hard work.