Allennlp Predictor class that mimics gec_model#6

Merged

ksteimel merged 25 commits intomasterfrom feature/predictor_from_gec_model

Oct 24, 2022

+5,444-10

Commits on Sep 28, 2022

words metadata is getting filled if unspecified when text_to_instance is called.
ksteimel
committed
Using JustSpacesWordSplitter so that tokenization matches that used by gec_model,
ksteimel
committed
Decode now adds the corrected sentence to the output dict.
ksteimel
committed
Updating gitignore to prevent adding .th files
ksteimel
committed
Adding directory fixture as analogue to model archive
ksteimel
committed
Fixing errors in modeling code now that model.decode adds the original words and corrected words to the output.
ksteimel
committed
Adding conditional so that no correction is performed in decode if no errors are identified.
ksteimel
committed
Appending start token when creating instances from json or string.
ksteimel
committed
Start token is expected in ouptut.
ksteimel
committed
Drop START_TOKEN from output_dict["words"]. This interferes with the offsets when creating the target (corrected sentence).
ksteimel
committed
The outputs now no longer have $START_TOKEN in the corrected sentence. This is identical to the output for gec_model!
ksteimel
committed
Handling multiple iterations of correction in predictor now.
ksteimel
committed
Changed location of weights file so it can be used by gec_predictor as well.
ksteimel
committed
setup is now downloading weights file if it does not already exist.
ksteimel
committed

Commits on Oct 12, 2022

Apply suggestions from code review

ksteimel
and
Frost45
authored
Removing unused imports, adding docstrings.
ksteimel
committed
Removing unused predictions to labeled_instances method.
ksteimel
committed
Updated docstring for decode()
ksteimel
committed
Removed unused imports.
ksteimel
committed

Commits on Oct 22, 2022

Adding back import of gec_predictor that shouldn't have been removed

ksteimel
authored and
ksteimel
committed