Unrecognized function or variable 'invertResidualLayer'

I have installed add-on

'Deep Learning Toolbox' and 'Deep Learning Toolbox model for mobilenetv2'

and used the below code

net = mobilenetv2; % Load pretrained MobileNet model
numClasses = numel(classNames);
layers = [
    net.Layers(1:end-3)
    fullyConnectedLayer(numClasses,'Name','fc')
    softmaxLayer('Name','softmax')
    classificationLayer('Name','classoutput')];

options = trainingOptions('adam', ...
    'MaxEpochs',10, ...
    'MiniBatchSize',32, ...
    'ValidationData',imdsValidation, ...
    'ValidationFrequency',5, ...
    'InitialLearnRate',1e-4, ...
    'LearnRateSchedule','piecewise', ...
    'LearnRateDropFactor',0.1, ...
    'LearnRateDropPeriod',5, ...
    'L2Regularization',0.01, ...
    'Plots','training-progress');

% Train the network
net = trainNetwork(imdsTrain,layers,options);

and I got the error message below

Error using trainNetwork
invalid network.

caused by:
    Layer 'block_11_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_12_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_14_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_15_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_2_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_4_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_5_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_7_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_8_add' Unconnected input. Each layer input must be connected to the output of 
another layer.
    Layer 'block_9_add' Unconnected input. Each layer input must be connected to the output of 
another layer.

I don`t know why this error happens...

maybe using pretrained model intrigue error depends on the data?

How to solve this error?

Answer 1

[Pratham Shah]

Hi,

I have struggled with this issue a lot and I know the reason behind it.

The problem is while you are calling 'net.Layers(1:end-3)' it will only have information of layers, not the connections and the problem is occuring in addition layer as it requires 2 inputs but the layer graph is skipping one connection in net.Layer.

I don't have any permenent solution to that but you can try 2 things:

1. Import the 'net' in the Network Designer Application and export the 'lgraph' after replacing last 3 layers with your desired layer.
2. You can use 'connectLayer' manually for all the unconnected layers like this,

lgraph = layerGraph(layers);

lgraph = connectLayers(lgraph,'input','add/in2');

but this process will take lot of time.

But by looking at you code I think you can just use the 'replaceLayer' in last layer(to match the number of classes) because mobilenetv2 already have fc and softmax function.

P.S. : Hope this helps :)

[NoYeah]

this helped me a lot but I won`t use MATLAB for deep learning