Run GLUE for fine-tuning Few-Shot Relation Classification about ask2transformers HOT 2 CLOSED

Hi!
Thanks for your comments I really appreciate them!

Regarding the question about the time required for training the model, yes, fine-tuning a model with this method takes time. In fact, it is proportional to the number of entailment/neutral/contradiction examples generated per TACRED example. I can give you some approximations based on my hardware:

GPU: 2080ti (11Gb)

With that configuration I am able to train roberta-large at 2.86 hours/epoch (whole dataset). Nevertheless, the results reported at the paper were obtained using posn and negn equal to 1, so, the training should be faster (check the paper Appendix B for more details).

Regarding the question about the output directory, also yes. Once you train your model you can pass the path to the pretrained_model argument to load and test your fine-tuned model. I was planning to upload already trained models to the HuggingFace Hub, but I am quite busy right now.

Let me know if you have any other question :)

from ask2transformers.

Thank you very much!

I actually found out that the problem was related to the --task_name mnli parameter, which was ignoring my local train/dev/test files. Removing this parameter made everything work as expected 🎉

from ask2transformers.