search

BICLab / SpikeYOLO

Offical implementation of "Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-performance and Energy-efficient Object Detection" (ECCV2024 Best Paper Candidate / Oral)
https://arxiv.org/abs/2407.20708
GNU Affero General Public License v3.0
110 stars 6 forks source link

There seem to be something wrong with power consumption calculation #25

Open Qiuyu-Li276 opened 1 day ago

Qiuyu-Li276 commented 1 day ago

Leaving aside the rest, as for MS_CnvBlock, there are a total of 3 convolutions in the RepConv of MS_CnvBlock, and it seems that the latter two have not been included in the calculation. Moreover, these latter two seem to be MAC operations, and the power consumption involved cannot be ignored. Why not consider them?

XinhaoLuo666 commented 1 day ago

As we mentioned in the paper, convolution here is a reparametric convolution, that is, it is reparameterized into a convolution and then computation is performed when it is re-inferred. Since reparameterization only needs to be performed once, there is only AC and no MAC for inference.

Qiuyu-Li276 commented 13 hours ago

As we mentioned in the paper, convolution here is a reparametric convolution, that is, it is reparameterized into a convolution and then computation is performed when it is re-inferred. Since reparameterization only needs to be performed once, there is only AC and no MAC for inference. This is the reparameterized convolution at the end of MS_ConvBlock, but there is a SepConv at the beginning, and there is a PWConv and a DWConv at the end of SepConv. There is no activation layer between the two conv, so DWConv should be counted as EAC, right? This doesn't seem to involve reparameterization operations in SepConv. image