[WIP] Lightning glue example #3290
Conversation
|
@srush Can you please take a look? |
Codecov Report
@@ Coverage Diff @@
## master #3290 +/- ##
==========================================
+ Coverage 77.81% 77.99% +0.18%
==========================================
Files 98 98
Lines 16666 16666
==========================================
+ Hits 12969 12999 +30
+ Misses 3697 3667 -30
Continue to review full report at Codecov.
|
srush
left a comment
srush
left a comment
There was a problem hiding this comment.
Nicely done. Pretty impressive how little code needed to change.
While you are doing this, I think you can be a bit more aggressive about moving code up to basetransformer and moving the manual logging to loggers. But otherwise the code looks great.
Not sure the best way to pull in BaseTransformer.py. The easiest is sys.path.append("../")
|
Looks excellent. I will let @LysandreJik merge tomorrow, and confirm multi-gpu / TPU work. Want to try SQuAD next? |
Sure, I'll give it a go. |
* ✨ Alter base pl transformer to use automodels * 🐛 Add batch size env variable to function call * 💄 Apply black code style from Makefile * 🚚 Move lightning base out of ner directory * ✨ Add lightning glue example * 💄 self * move _feature_file to base class * ✨ Move eval logging to custom callback * 💄 Apply black code style * 🐛 Add parent to pythonpath, remove copy command * 🐛 Add missing max_length kwarg
4 participants
This PR adds an example of using Pytorch Lightning to run the GLUE benchmark. Additionally, I altered the
transformer_base.pyto use auto models and moved it to the example directory so it could be copied in by any script that wishes to use it.Preferably, the base transformer would have subclasses for the different types of tasks, but I just used a dictionary with a key passed on init instead. (i.e. NER uses
AutoModelForTokenClassificationand GLUE usesAutoModelForSequenceClassification).