Open josephzpng opened 3 months ago
Good question!
You are right! Thank you for pointing out this misalignment. We attempt lots of designs for generating condition embeddings, as shown in Table 5 in paper. There could be a misalignment when I refactor this code and choose a version with avg_pooling and corresponding ckpt. I will update the paper version asap. And before that, you can firstly follow current code (both are avg_pooling) if you are doing some finetuning on current PSALM ckpt, since current version also got a great performance when I evaluate on it :)
Thanks for your interest in our work!
Great Job! Regarding the Ref task, the paper says that the Sentence Condition is extracted from the special [REF] token, while the embedding of Sentence Condition in the code is obtained by avg_pooling the entire sentence, so should I follow the code?