Why did we freeze these packages?

[lgessler]

The following packages are frozen in requirements.txt:

flair==0.4.5
pandas<0.24
scikit-learn==0.22.1
torch<1.6
xgboost==0.90

[lgessler]

torch<1.6: this is because we're still using stanfordnlp, which uses some code which uses a deprecated matrix operation. (Upgrading stanfordnlp to stanza will fix this--see #3)
scikit-learn: there are pickles somewhere in our repo that we're loading that were produced under this scikit-learn version
flair==0.4.5: various pieces of code make assumptions about Flair's API and semantics that are no longer true for flair>0.4

[nitinvwaran]

I was involved in the freezing of xgboost. The POS ensemble model would not load for xgboost > 0.90 because of serialization issues. Don't recall the exact error but i recall v1.0 did upgrades to their serialization mechanism that weren't backward compatible. More details are here: https://xgboost.readthedocs.io/en/latest/tutorials/saving_model.html