Closed valencebond closed 9 months ago
Hi, due to the internship change, I have no access to the previous log. But reproducing Maskalign is very easy, so my suggestion is first pre-training on ImageNet and then checking the loss status. You can also discuss more details with me.
I use your code to pretrain video datasets, But I'm not sure what is the normal rate of loss decline