Group Layers in Deep Network Designer

I am playing a bit around with the Deep Network Designer and struggle with huge architectures like YOLOV5FACE. The biggest constraint is the C3 Layer which includes itself 27 Layer. Even if I export parts of the network and import them back from WS the result is a constant growing structure with little to none overview. Maybe I am a bit spoiled from the subsystem function from simulink but in my opinion the Deep Network Designer lags this function or I am not able to find it. I only see at the moment to create a custom layer for each subsitute network structura as a solution.

Answer 1

[Philip Brown]

It's not currently possible to view nested "subsystem"-like blocks in Deep Network Designer. We're aware this would be a useful feature to support, especially for larger and more complex network architectures.

As you mentioned, at the command line you could make a custom layer containing the network, like in this example.

[Matti Kaupenjohann]

This example really helps, didn't came up with name nested, so I had problems with using the predict method. I was thinking, I need to use the predict method of the inner layer, which I realized is not really reachable.

[Matti Kaupenjohann]

I worked through the example and created a custom CBS block and recognized some weird behaviour on the dlnetwork function. The function returns:

Error using dlnetwork
Too many input arguments.

Error in CBSLayer (line 62)
            dlnet = dlnetwork(lgraph,'Initialize',false);

So I debugged the class and used the function without the 'Initialize' and got back:

Error using dlnetwork (line 111)
Invalid network.

Caused by:
    Network: Missing input layer. The network must have at least one input layer.
    Layer 'conv1': Unconnected input. Each layer input must be connected to the output of another layer.
    

So I copy&paste the complete example and the example is not happy about the missing Normalization class:

    Unrecognized function or variable 'groupNormalizationLayer'.

Error in residualBlockLayer (line 59)
                groupNormalizationLayer('all-channels','Name','gn1')

Is the Example up-to-date?