peter-sk
peter-sk
@ArthurZucker Sure. I'm getting the hang of it. Now, the only failing tests are connected to flax and seem unrelated to this pull request.
If the flax errors are not due to the PR, this is ready to be reviewed, @ArthurZucker and @younesbelkada :-)
I just checked the logs for the remaining errors one more time. The errors are related to the import of the optax library, where jax.Array is used in a type....
The jax version used in the examples_flax test is 0.3.6: Collecting jax!=0.3.2,=0.2.8 (from transformers==4.28.0.dev0) Using cached jax-0.3.6-py3-none-any.whl This version clearly has no Array class. I am unsure why such an...
Figured out that optax
Definitely ready for review, @ArthurZucker and @younesbelkada :-)
All done and ready to be merged, @ArthurZucker and @younesbelkada 👍
I implemented the same change as for GPTNeoXForTokenClassification, i.e., I removed the hasattr etc. and just use config.classifier_dropout directly.
@sgugger Ready to merge when the checks complete. Thanks for the fast action 👍 ... and more to come in the next weeks!
I am doing this model on all languages except for NER tasks.