EfficinetNetV2 uses MBConv/Fused-MBConv as a part of it's architecture. There is no clarity of what these operations actually are from the paper (nor from the references). It appears that it is some sort of soft switch from Depthwise convs to normal convolutional layers, which would depend on the actual time to execute on the accelerator. Is there a definition of MBConv/Fused-MBConv anywhere?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
3
The bottleneck_block used as the basic building block of MobileNetv2 is the MBConv (building block of EfficientNets). Please refer bottle_neck method in https://towardsdatascience.com/mobilenetv2-inverted-residuals-and-linear-bottlenecks-8a4362f4ffd5
MBConvs seek inverted Res nature, which is the exact contrary to the working of ResNets.. It should all be clear in the above link
It can be seen in many papers involving EfficientNets, that MBConvs are basically the structure of MobileNetv2
-
1$\begingroup$ In efficientNet (and efficientNet v2) the MBconv used mention a layer called "SE". (see arxiv.org/pdf/2104.00298.pdf, figure 2). I dont see this layer mentionned on your linked article and I can't find any sources on it. Could you help ? $\endgroup$– ChopinMay 24, 2021 at 9:16
-
1
-
$\begingroup$ I found the explanation in this article more clear. towardsdatascience.com/… $\endgroup$ Jul 13 at 17:29