How to transform fully connected layer into convolution layer?

Question

QU HAIFENG 2018-6-11

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/405001-how-to-transform-fully-connected-layer-into-convolution-layer

编辑： Matt J 2022-4-27

I want to replace a fully connected layer with a convolution layer, but I don't know how to complete it. I reference help document, and find a function named "transposedConv2dLayer", but I don't believe "cropping" is helpful. So if somebody could give me any suggest, thanks a lot.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Shounak Mitra 2018-6-21

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/405001-how-to-transform-fully-connected-layer-into-convolution-layer#answer_325680

编辑：Matt J 2022-4-27

Hello QU,

Can you please tell me if you're using a Series or DAG network? The steps to replace a layer is different for these 2 network types.

For a Series Network, follow steps here -->

https://www.mathworks.com/help/nnet/examples/transfer-learning-using-alexnet.html#d119e3782

For DAG follow this -->

https://www.mathworks.com/help/nnet/ref/nnet.cnn.layergraph.html?s_tid=doc_ta#d119e84822

Also, is the motivation behind replacing the FC layer with a conv2d is to make the network deeper?

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Matt J 2022-4-27

编辑：Matt J 2022-4-27

The link for the Series Network does not seem applicable. The fully connected layer in that example is replaced by another fully connected layer, not a conv2d layer. Moreover, the new layer is not equivalent to the old one.

请先登录，再进行评论。