Closed windhxs closed 2 years ago
Hi @windhxs, the final step of PET training is similar to knowledge distillation where the predictions of different models trained with different patterns are combined and a final model is trained on these combined predictions. The logits
field in InputExample
is used to store the combined (i.e., averaged) per-class logits to train the final model. This happens in line 254 - 261 of modeling.py
.
Accordingly, the original InputExample
s that you provide to the model should not have any logits
.
That's for answering!
In class
InputExample
, it has a propertylogits
, I am confused of this property, what does it used for ?