ucinlp / null-prompts Goto Github PK

Adding new parameters to experiments can break existing crossval configs, since it is assumed that all parameters are exhaustively listed. It would be nice if crossval could assume default parameters to avoid this issue.

Hi !
Thanks a lot for your nice work and outsourcing the code.

I arrived a bit late on the poster presentation !

Which version of GPU + memory is required to run your code ? One LM per gpu or several gpu per LM ? This does not seem to appear in the paper version :)

Thanks a lot for your reply.

The code added to support priming introduced a couple bugs:

1. PyTorch doesn't want to backprop through repeated trigger embeddings (necc. to combine priming with learned prompts).
2. Some kind of memory leak or large memory requirement.