-
Notifications
You must be signed in to change notification settings - Fork 8
Previous Updates (February 2019)
Bahadir Sahin edited this page Mar 2, 2019
·
1 revision
- Code development for NER evaluation (F1 scores for tag based, bio+tag based) still continues. I slowed down it a little bit due to personal life issues.
- Decided to simplify README.md and use Wiki of this repository.
- All entries related to the updates in January 2019 are moved to related Wiki page.
- I will save my experiment results in Wiki, too.
- While trying to figure out how to use README and Wiki more efficiently, I will figure better things out hopefully =)
- As you may noticed, TextCNN experiments are finished. I will continue with LSTM/GRU experiments (first, I have to figure out a good parameter set =)).
- Since I separated classification and NER trainers/evaluators, I decided to create "scorer" folder to prevent bloating "utils.py" with metric calculation functions.
- "scorer/" will contain current and future metric calculation methods (not giving too much details, since you can always check it =)).
- I encountered some bugs in NER training and evaluation processes due to save/load functionalities. Hopefully, I fully fixed them. but if anyone out there reading this and using this repository, if you find any bugs, just let me know.
- Made some minor changes in namings and indexing (not much crucial stuff, details can be found in git commit message).
- Personal life issues still continues, hence slow development-slow experiment mode still continues.
- Precision, recall and F1 metrics are added into "ner_scorer.py".
- Since these metrics must be calculated for full set (not batch-based), I changed the evaluator flow a little bit.
- Evaluator reports mean precision, recall and F1 scores over all tags/named-entities.
- Detailed, tag-based, scores can be also reported by activating boolean detailed_ner_log (default value is true).
- In LSTM, I encountered a minor bug while using "bidirectional=true". Hopefully, it is fixed (at least training/evaluating was working on a small set).
- I tried a larger set to see whether my code is working but I got a "cuda illegal memory access" error. I think it is because OOM issues, but I am not sure for now.
- An "allowed_transition" stuff will be added in near future (like allennlp/conditional_random_field.py).
- Also, I updated my libraries, hence requirements.txt is changed, too =)
- Personal life issues still continues, hence slow development-slow experiment mode still continues.