Bug in F1 score computation

[trangham283]

Hello, thank you for releasing the code!

I'm wondering if this line for computing the F1 score is a bug (flipped order of truth and pred), or it's because of how your data is set up. If so, could you elaborate?

A similar issue is raised here.

[martinmamql]

Hi Trang,

Greetings! Thank you very much for pointing it out -- it is indeed a bug of evaluating the F1 score in MOSEI sentiment and I have corrected the code. Using the correct version, I can obtain an F1 score of over 82% within 20 epochs on MOSEI sentiment, using the default run.sh in the Github page, which is even better than the reported number in the paper (81.8%). More hyper-parameter tuning might help the performance.

Thank you very much for your post!

Sincerely,
Martin