BatchNorm consuming a lot of time

[leochli]

Hi shicai,

Have you evaluated the time performance of mobilenetv1 & v2?
I found that the bn layer and linear layer are very time consuming both in CPU-only and GPU-only cases. Is the eps setting really necessary for the BN layer?

[basaltzhang]

Try to use the DepthwiseConvolution layer instead of Convolution type, if the group parameter is not 1. The DepthwiseConvolution implementation can be found here

[AnPay]

Can someone update the time performance for mobilenets?
i am getting it ~300ms for googlenet. Want to verify if its correct. tested on S10