Are dlnetworks supposed to be allowed to have output layers?

Are dlnetworks allowed to have output layers? In the following code, I manage to create one, so the answer would seem to be yes.

 layers= [imageInputLayer([1,1,1])  , reluLayer(Name='relu') ] ;

 dln = replaceLayer( dlnetwork(layers) ,'relu', regressionLayer);
 
 class(dln)

ans = 'dlnetwork'

 dln.Layers

ans = 
  2x1 Layer array with layers:

     1   'imageinput'         Image Input         1x1x1 images with 'zerocenter' normalization
     2   'regressionoutput'   Regression Output   mean-squared-error

However, when I try to create this more directly, an error is raised:

 dln = dlnetwork(  [imageInputLayer([1,1,1])  , regressionLayer] )

Error using dlnetwork/initialize (line 600)
Invalid network.

Error in dlnetwork (line 182)
net = initialize(net, dlX{:});

Caused by:
Layer 'regressionoutput': Detected output layer. The network must not have output layers.

Have I found an unintended backdoor?

Answer 1

[Jack]

By design, dlnetwork objects are intended for custom training loops and are not supposed to include output layers like regressionLayer or classificationLayer. If you try to create a dlnetwork directly with an output layer, MATLAB throws an error. The fact that replaceLayer can slip in an output layer is effectively a workaround, but it isn’t officially supported.

Follow me so you can message me anytime with future questions. If this helps, please accept the answer and upvote it as well.

[Matt J]

Hi @Jack. Is that documented somewhere? I don't actually see a doc passage saying output layers are forbiddden.

[Jack]

Regarding documentation, while you might not find an explicit statement saying “dlnetwork objects must not contain output layers,” the behavior and error messages in MATLAB reflect this design choice. The documentation and examples for dlnetwork consistently show networks built without an output layer, reinforcing that output layers are meant to be handled externally in your training loop.

[Jack]

@Matt J did this clear it up?