Future Model Logistics and Metadata

[bollwyvl]

Hi there! I'm interested in writing (and maintaining, though always appreciate having upstream maintainers around, if just to keep me honest) a conda-forge recipe for neuralcoref. I've started some thoughts here, and think we can offer users some pretty nice ergonomics around installing/using the package without much on-going effort on this side of the house (congrats on the py2 spring cleaning)!

I'm presently chugging along, dealing with layering in the spacy/**/*.pxd during build, and fighting with window (always), but it would be helpful for downstreams if a future version of neuralcoref did a few, relatively small things:

• introduce additional search locations for NEURALCOREF_CACHE
  • this would allow us to package neuralcoref-model, and save you some s3 :moneybag:
  • e.g. os.path.join(sys.prefix, "share", "neuralcoref_cache")
  • if i can get a 4.0 building in a reasonable amount of hours, I'll probably patch to just do this
  • allow setting something at runtime
  • at worst, allow an environment variable
  • not preferred, as people don't always activate their environments fully
• add a version string (and publish SHAs) of the model on s3
• include a LICENSE in the model tarball
  • I've no opinion as to what would be appropriate: the spacy models are MIT, I suppose
  • they are also pip installable, which is nice
• at least bottom pin the package versions needed in train
  • we can set some runtime pins to make sure we don't get something totally crazy
  • or, while lacking pip-style extras, we can offer a neuralcoref-train package which includes torch et. al

Happy to PR any of the above, wanted to document things as I discover gotchas!

[bollwyvl]

at worst, allow an environment variable

Hah, did find the environment variable, and have been able to make it be found with some activation scripts. However, running into issues finding a version of conda-forge spacy + sources that will build against the hello world example...

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.