hustvl / Vim

Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Mamba module initialization

mingtan2 opened this issue · comments

Dear authors,

Thanks for the great work and releasing the codebase! I have one question regarding the Mamba module initialization.

From the code, I see its linear layers are initialized here and here, then the weights and biases are overwritten here and here, which zeros out the mamba linear layers' biases. So this line actually can be deleted? Please share your insights if it's intended. Thanks!