Visualising the features in a convolutional neural network

I’ve been using the matlab example for visualisation of features of a CNN model and visualized the convolutional network layers of all the sequential DL models. With the R2017b release we have the DAG network googLeNet as well. I’m not able to use the available code for sequential models to visualize the features in the googLeNet DAG model. Do we have a way of doing this? Can matlab give me an example for this visualisation?

[Kenta]

Yes, you can do that with matlab. For example, please refer to the following link

https://jp.mathworks.com/matlabcentral/fileexchange/74639-continuous-activation-map-generation-at-training-process

Answer 1

[Javier Pinzón]

Hello All

Activations can be see in a DAGNetwork, only the way of write it variesfor the DAG, You need to select the Layer which want to get the activations, and then follow the next lines:

[height1, width1, ~] = size(im);

 if isa(net,'DAGNetwork') == 1    
    featureMap = activations(net, im, 'Relu_conv1');
    % The feature map is stored in the first channel
    Map1 = featureMap(:, :, 1);
    Map1 = imresize(Map1, [height1, width1]);

 figure
    montage(cat(4,im,...
        repmat(Map1*255,[1 1 3])), 'Size', [1 2]);
 end

the activation feature map is in the position 1 of the matrix, and you can do that with any of the layers.

Hope it helps and is what you wanted. If any question, you may ask me.

Regards,

Javier

[Muhammad Arsalan]

Hello,

How can we find the feature like in attached image from a DAGNetwork. The code you provided does not give the color features as attached in attached image by deeDreamImage. or is there any way to use deepDreamImage for DAGNetwork ???

[Salma Hassan]

i confused with this problem when i visualize the feature by the code in this link https://www.mathworks.com/help/nnet/examples/visualize-features-of-a-convolutional-neural-network.html i get features for alexnet architecture , but I expected features for my dataset which is completely different than alexnet

Answer 2

[keke zhang]

Hi Sivaramakrishnan Rajaraman: I also meet the problem, are you solve it now?

[Sivaramakrishnan Rajaraman]

Hi, The visualization option is available only for the sequential models and not for the DAG models. We need to wait for an update from Matlab :)Did you find any other option?

[keke zhang]

Hi,I am not find other options. It is a pity.

[Arun A M]

https://in.mathworks.com/help/deeplearning/ug/visualize-features-of-a-convolutional-neural-network.html

Answer 3

[Philip Brown]

If you're looking for using the deepDreamImage function for DAGNetwork networks, that's supported from R2019b.