NVlabs/SegFormer

k v compression in attention may causes small targets and detail to be lost？

Opened this issue · 0 comments

swjtulinxi commented

In attention， to reduce the calculation amount, kv is compressed， small targets and detail will lose ？