Is it possible to view the features of a dagnet network?

Question

Ashley 2018-8-30

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/416910-is-it-possible-to-view-the-features-of-a-dagnet-network

回答： Philip Brown 2023-5-4

I'm trying to create examples of images that 'strongly activate' channels of convolutional layers of a CNN (see https://www.mathworks.com/help/nnet/examples/visualize-features-of-a-convolutional-neural-network.html).

I've tried using the function deepDreamImage() but it only works for SeriesNetwork types, which I cannot use because my network has filters in parallel. Is there an equivalent function or workaround I can use for a DAGnetwork?

act1 = activations(net,im,'conv1','OutputAs','channels')

works on my network but produces greyscale outputs and I'm looking for color images.

I tried:

layer=5;

channels=1:8;

I = deepDreamImage(net,layer,channels,'PyramidLevels',1);

MatLab Error: Error using deepDreamImage Expected input number 1, network, to be one of these types: SeriesNetwork Instead its type was DAGNetwork.

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

Elif BAYKAL 2018-11-21

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/416910-is-it-possible-to-view-the-features-of-a-dagnet-network#answer_348162

layer = 2;

name = net.Layers(layer).Name

net = SeriesNetwork(net.Layers);

channels = 1:56;

I = deepDreamImage(net,layer,channels,'PyramidLevels',1);

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Stuart Whipp 2018-12-10

Does the above answer yorur question? I've developed a function for activations of recurrent networks such as LSTM. This could possibly be amended to work for DAGs, though in my experience the DAG layers are listed at the end, after output layers (in terms of layer index).

You'd want to ensure you were removing only the layers after the one you're interested in (in terms of layer architecture sequence not index) and I'm guessing the layer selected would have to follow an addition/depth concat layer otherwise there'd be two parallel/disparate outputs? In essence, this method applies a regression layer at the end of a sliced network and so will only output from one layer as though it was the penultimate layer of a net. Hope that makes sense!

LSTM activations

请先登录，再进行评论。