I'd like to ask how the Uniad model achieves lightweight optimization.

[hehaibo-ok]

I tried cutting out two layers of the codec and found that its score on the dataset plummeted dramatically.