How can I modify the weights of a pre-trained Alexnet in 2018a?

Question

Valerio Gloria 2018-6-22

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/406918-how-can-i-modify-the-weights-of-a-pre-trained-alexnet-in-2018a

回答： Mingrun Wang 2018-6-22

Hi everybody. For a research project I need to modify the weights of a pre-trained Alexnet. When I try to do so:

net = alexnet;
net.Layers(2).Weights = net.Layers(2).Weights * 0.01;

Matlab 2017b answers me that they are read-only values. Since it looks a very strong limitation, does anybody know if in version 2018a they solved that problem? Or maybe someone has solve it with a script?

Thanks.

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

Mingrun Wang 2018-6-22

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/406918-how-can-i-modify-the-weights-of-a-pre-trained-alexnet-in-2018a#answer_325823

layers=net.Layers; layers(2).Weights=layers(2).Weights * 0.01; And i also provide you the following ways to make up for the conversion between external layer and internal layer.(external layer is the wrapper for internal layer) externalLayer→internalLayer: nnet.cnn.layer.Layer.getInternalLayers( externalLayers ) %this function will return an internal instance(copy).

internalLayer→externalLayer:externalLayersConstructor(e.gnnet.cnn.layer.FullyConnectedLayer(internalLayer));

create an internal layer(e.g nnet.internal.cnn.layer.FullyConnected(name,inputsize,outputsize));