caffe-mobilenet-v3 icon indicating copy to clipboard operation
caffe-mobilenet-v3 copied to clipboard

about Squeeze-And-Excite in that block

Open zys1994 opened this issue 6 years ago • 5 comments

I Notice that you don't add Squeeze-And-Excite in some block which is different from paper. why not add it?

zys1994 avatar May 17 '19 06:05 zys1994

I Notice that you don't add Squeeze-And-Excite in some block which is different from paper. why not add it?

I tried to implement the network architecture as described in the paper. Please specify the location of the problem you mentioned, I will check it.

jixing0415 avatar May 17 '19 12:05 jixing0415

怎么训练

zxx741X avatar May 28 '19 15:05 zxx741X

Did you reproduce the paper's acc?

runningJ avatar May 30 '19 09:05 runningJ

Did you reproduce the paper's acc?

Not yet, I just got 65% accuracy on v3_small and 70% accuracy on v3_large. I think there may be some training strategies that are not in the way mentioned in the paper.

jixing0415 avatar May 31 '19 06:05 jixing0415

hello, i want to kown why layer"conv2-1/dwise/scale" end with layer"conv2-1/dwise/scale-relu", while other "se" layer only have layer "scale"?

jiqirenno1 avatar Jul 17 '19 10:07 jiqirenno1