Question about details of Task Branch #8
Description
Hi, thank you for sharing such an awesome project. I have some questions about the details in your code, and I sincerely hope you can help me solve them.
1 In your paper, it is said that "We conduct knowledge amalgamation for each block of TargetNet" (Section 5.1.1), however, in the code, it seems that you only apply the knowledge amalgamation on the decoder blocks. This makes me feel confused that whether we need to apply it on the encoder blocks, as in the "layerwise amal".
2 In the task branch code, should the object of student be JointSegNet, instead of BranchySegNet? As in BranchySegNet, the network does not use teacher's blocks.
Since after running the code on NYUv2, I only got a target net with mIOU 0.2830 by merging the depth model (RMSE: 0.6772) and the seg model (MIOU: 0.5227) with BranchySegNet. When I replace it with JointSegNet, I can get a target net with mIOU 0.5233.
I would really appreciate it if you could help me. Look forward to your reply, thanks!