Closed Dwrety closed 2 years ago
Hi, the embed_extractor
is used to ensure the same channel of the generated kernel and that in the encoded feature.
Because the channel in encoded feature may change in different settings. So, the embed_extractor
is adopted for adjustment.
I see, so it is basically a linear projection. Thanks for your clarification.
Is it just a linear project layer? After acquiring the kernel weights from the KernelHead, I cannot find any activation layer following the layer itself. The weight seems to go directly to another Conv2d layer (embed_extractor) without any non-linearlity. If so, why not directly producing the kernels in the kernelhead instead?