Can I use the "activation" function to get multiple layer activations from a "SeriesNetwork" with one function call in R2022b?

I would like to obtain activations for multiple layers in a "SeriesNetwork" using "activations". However, "activations" only allows for specifying one desired layer at a time. Therefore, I have to call "activations" multiple times, which can be slow and result in duplicated computations. Is there a faster way to do this?

net = alexnet; % A SeriesNetwork
img = rand(227,227,3); % Random image
ac2 = activations(net, img, 2); 
ac3 = activations(net, img, 3); 
ac4 = activations(net, img, 4); 
ac5 = activations(net, img, 5);
% ...

Answer 1

[MathWorks Support Team]

As of R2022b, it is not currently possible to use the "activations" function to compute activations for more than one layer in the same function call. However, the development team is aware of this limitation and may consider addressing it in the future releases.

As a workaround, you could convert the “SeriesNetwork” to a “dlnetwork”. The “dlnetwork” type supports computing activations in multiple layers using “predict”. Below is an example of this workaround: 

net = alexnet; % SeriesNetwork 
% Convert to dlnetwork 
lgraph = layerGraph(net); 
lgraph = removeLayers(lgraph,["prob" "output"]); % To convert to dlnetwork we need to remove the output layer 
dlnet = dlnetwork(lgraph); 

% Prepare data 
img = rand(227,227,3); % Random image 
img = dlarray(img, 'SSCB') % “dlnetwork.predict” requires a formatted “dlarray” as input 

% Get activations 
outputActivationLayers = ["conv1", "conv2", "fc6"]; % “dlnetwork.predict” requires specifying desired layers using strings 
[ac1, ac2, ac3] = dlnet.predict(img, Outputs=outputActivationLayers); 

Though this workaround requires some conversions of both network and data, it should be faster than calling “activations” numerous times. Note that if you would like to switch to working with "dlnetwork" entirely, to train a “dlnetwork” you need to write a custom training loop.