Code Question

[Ahnsun]

DNL-Object-Detection/mmdet/ops/nl_gc_block.py

Line 151 in 6ae8884

out_sim = torch.bmm(sim_map, value.transpose(1, 2))

Why do the matrix multiplication V×(Q×K) before adding Wm?

[Howal]

It follows Eq.12 in our paper.

[Ahnsun]

Thanks for your answer,but in the Fig. 2，Wv is the last one to be multiplied after adding the Wm,and in the code,value(Wv) is multiplied before adding the Wm

[Howal]

Vx(QxK+Wm) = Vx(QxK) + VxWm
We follow the left term in Fig 2 and the right term in the code.