Error Network: Missing output layer

I am creating an autoencoder where I want to use the Alexnet network for the encoder part (removing the last layers), and when I try to train the autoencoder, I get the error:

"Error using trainNetwork

Invalid network.

Caused by:

Network: Missing output layer. The network must have at least one output layer.

Layer 'output': Unconnected output. Each layer output must be connected to the input of another layer."

However, I do see the output layer of the decoder connected when I visualize the network graph before training.

Please, your help on how to solve the problem. Thank you.

[Debraj Maji]

Hi @Fernando Bonilla Hidrobo,

Can you please share the code for the network?

[Fernando Bonilla Hidrobo]

Hi @Debraj Maji, this is the code of the autoencoder I am creating

alexNet = alexnet;
lgraph = layerGraph(alexNet.Layers(1:end-3)); % Remove the last three layers

bottleneckLayer = fullyConnectedLayer(256, 'Name', 'bottleneck');
lgraph2 = addLayers(lgraph, bottleneckLayer);

% Get the name of the last layer of the modified encoder
lastEncoderLayer = lgraph.Layers(end).Name;
lgraph2 = connectLayers(lgraph2, lastEncoderLayer, 'bottleneck');

classInput = imageInputLayer([1, 1, 4], 'Name', 'classInput', 'Normalization', 'none');
concatLayer = concatenationLayer(3, 2, 'Name', 'concat');
lgraph3 = addLayers(lgraph2, concatLayer);
lgraph4 = connectLayers(lgraph3, 'bottleneck', 'concat/in1');
lgraph5 = addLayers(lgraph4, classInput);
lgraph6 = connectLayers(lgraph5, 'classInput', 'concat/in2');

analyzeNetwork(lgraph8);

%%  decoder
outputImageSize = [227, 227, 3];% (width x height x channels)

decoderLayers = [
    transposedConv2dLayer(3, 64, 'Stride', 2, 'Cropping', 'same', 'Name', 'decoder_conv1')
    reluLayer('Name', 'decoder_relu1')
    transposedConv2dLayer(3, outputImageSize(3), 'Stride', 2, 'Cropping', 'same', 'Name', 'decoder_conv2')
    ];
outputLayer = convolution2dLayer(1, outputImageSize(3), 'Name', 'output');
lgraph7 = addLayers(lgraph6, decoderLayers);
lgraph8 = connectLayers(lgraph7, 'concat', 'decoder_conv1');
lgraph8 = addLayers(lgraph8, outputLayer);
lgraph8 = connectLayers(lgraph8, 'decoder_conv2', 'output');

analyzeNetwork(lgraph8);

%% Image Loading
datafolder = 'C:\Users\ferna\Desktop\U\10TH SEMESTER\MIC\IMAGES\FFT_2048\bicubic';
imds = imageDatastore(datafolder, 'IncludeSubfolders', true, 'LabelSource', 'foldernames');
[trainingData, validationData] = splitEachLabel(imds, 0.8, 'randomized');
options = trainingOptions('adam', ...
    'MaxEpochs', 20, ...
    'InitialLearnRate', 0.0001, ...
    'ValidationData', validationData, ...
    'Plots', 'training-progress');
autoencoder = trainNetwork(trainingData, lgraph8, options);

[Cris LaPierre]

What do you want the output of your network to be? The output options can be viewed here:

https://www.mathworks.com/help/deeplearning/ug/list-of-deep-learning-layers.html#mw_d62e1731-91aa-4182-80b1-c9f5c3cb4aa9

Answer 1

[Matt J]

The "output layer" referred to by the error message doesn't refer to the final decoder in the network. An output layer is a specific type of layers that implements a loss function for the purpose of training,

https://www.mathworks.com/help/deeplearning/ug/list-of-deep-learning-layers.html#mw_d62e1731-91aa-4182-80b1-c9f5c3cb4aa9

You must have one of these as the final layer in your network, so that trainNetwork knows what loss function to use.