search

ChengpengChen / RepGhost

RepGhost: A Hardware-Efficient Ghost Module via Re-parameterization
MIT License
168 stars 17 forks source link

The latency on the snapdragon 8 gen 1 is almost the same before and after re-parameterized #6

Closed neptune4year closed 1 year ago

neptune4year commented 1 year ago

Thanks very much for sharing this excellent work. Our whole model is an U-net style model with RepGhost as the backbone, and we convert it to ONNX and then convert it to MNN. But we found that the latency is almost the same before re-parameterized (15ms) and after re-parameterized (13ms)， would it be correct?

ChengpengChen commented 1 year ago

Hi, neptune4year,

The speedup (13.3%) is to be expected, because the re-parameterization structure (only one BN for dw conv) of our network is much simpler, compared to those of RepVGG and MobileOne.

Beside, thanks for extending our network to a U-Net style. Please feel free to contact me, If you have any questions about its application.

neptune4year commented 1 year ago

Thanks a lot, I got it.