I’m fintuning T5 for superglue datasets.
Now, I use first half of ‘validation’ set as validation set and the second half of ‘validation’ set as test set.
I’m little bit confused of how to pass decoder_input_ids
on validation step.
For example,
input_ids
, attention_mask
could be obtained from tokenize("My model outperformed baseline models")
.
And in my opinion, I should give just start token as decoder_input_ids
with the variables above to forward()
. Is it correct?
This results in the length of the output become 1(same as start token).
Do I need to just pass tokenized ‘labels’ without decoder_input_ids
like in training step?
And I’m not sure if I need to use generate()
instead of forward()
.