Question about Task 3 #4
Description
Dear Author:
Congratulations for your excellent work! I have some questions:
1) In task 3, you use long term model pretrained in task 2 to align features but we don't froze long term features from task 2 in task 3 which means long term features may still be trained in aligning progress in task 3.
2) In PPT, you use causal attention.What's the difference between mask type 'causal' and 'all'?
3) I noticed in task 2, you use warm-up. What about the experiment result without this progress in sdd dataset?
Looking forward for your reply!