Open Little-Podi opened 1 year ago
I have the same question about the MHA in this paper. Have you solved this problem yet ?
As far as I know, there is no official complete code currently open-sourced online. You need to implement the remaining functions according to the paper by yourself.
Thanks for your excellent work and wonderful open-source code! When I read the code, I found a possible issue about the implementation of MHA message fusion part. In this line (https://github.com/MediaBrain-SJTU/Where2comm/blob/efeee4a33b78c362b277d7f698ad4b3b2da0392c/opencood/models/fuse_modules/where2comm_attn.py#L100), you feed the spatial confidence map as an input of parameter
I think the three terms of input parameters correspond to the query, key and value of the multi-head attention mechanism, respectively. My confusion is, since the output of this formulation is the attetion weight, why we need the value to serve as an input? Maybe you express in this form for the sake of generality, so my main concern is about the code implementation. Thanks a lot for your reading. Looking forward to your reply!
quality_map
. However, in this line (https://github.com/MediaBrain-SJTU/Where2comm/blob/efeee4a33b78c362b277d7f698ad4b3b2da0392c/opencood/models/fuse_modules/where2comm_attn.py#L76), theself.attn
is implemented by the official classnn.MultiheadAttention
, which doesn't have a parameter that corresponds toquality_map
. Thus, it seems that the code cannot run smoothly. Maybe I miss something. Could you please give me some guides? Another minor question is, for the equation (3) in your paper