connecting concenation layer error

Hello everyone. I have an issue. In the following code, I cant connect concatenationLayer = concat to featureAttention & temporalAttention. Would you please help?

Error Message

Caused by:

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

++

numFeatures = size(XTrain, 2);
numClasses = numel(categories(YTrain));
% Feature-Level Attention
featureAttention = [
    fullyConnectedLayer(64, 'Name', 'fc_feature_attention')
    reluLayer('Name', 'relu_feature_attention')
    fullyConnectedLayer(1, 'Name', 'fc_feature_weights')
    softmaxLayer('Name', 'feature_attention_weights')
    ];
% Temporal Attention (not used for Iris dataset, but included for completeness)
temporalAttention = [
    fullyConnectedLayer(numFeatures, 'Name', 'input_sequence')
    lstmLayer(64, 'OutputMode', 'sequence', 'Name', 'lstm_temporal_attention')
    fullyConnectedLayer(1, 'Name', 'fc_temporal_weights')
    softmaxLayer('Name', 'temporal_attention_weights')
    ];
% Combine into Hierarchical Attention
hierarchicalAttention = [
    featureInputLayer(numFeatures, 'Name', 'input_features') % Input layer for features
    featureAttention
    temporalAttention
    concatenationLayer(1, 2, 'Name', 'concat') % Concatenate feature and temporal attention outputs
    ];

[Matt J]

We cannot demonstrate solutions, because you do not provide numFeatures or numClasses. They are generated from XTrain and YTrain, which we do not have.

Answer 1

[Matt J]

Use connectLayers to make your connections programmatically or make the connections manually in the deepNetworkDesigner.

[Matt J]

If you have R2024a, you can simplify things a bit with networkLayer:

% Feature-Level Attention Block (Encapsulated)
featureAttention = networkLayer([
    fullyConnectedLayer(64, 'Name', 'fc_feature_attention')
    reluLayer('Name', 'relu_feature_attention')
    fullyConnectedLayer(1, 'Name', 'fc_feature_weights')
    softmaxLayer('Name', 'feature_attention_weights')
    ], 'Name', 'feature_attention_block');

% Temporal Attention Block (Encapsulated)
temporalAttention = networkLayer([
    fullyConnectedLayer(numFeatures, 'Name', 'input_sequence')
    lstmLayer(64, 'OutputMode', 'sequence', 'Name', 'lstm_temporal_attention')
    fullyConnectedLayer(1, 'Name', 'fc_temporal_weights')
    softmaxLayer('Name', 'temporal_attention_weights')
    ], 'Name', 'temporal_attention_block');


% Create a layerGraph with multiple layers but NO connections yet
hierarchicalAttention = layerGraph([
    featureInputLayer(numFeatures, 'Name', 'input_features');
    featureAttention
    temporalAttention
    concatenationLayer(1, 2, 'Name', 'concat_attention');
]);

% Connect the layers
hierarchicalAttention = connectLayers(hierarchicalAttention, 'input_features', 'feature_attention_block');
hierarchicalAttention = connectLayers(hierarchicalAttention, 'input_features', 'temporal_attention_block');
hierarchicalAttention = connectLayers(hierarchicalAttention, 'feature_attention_block', 'concat/in1');
hierarchicalAttention = connectLayers(hierarchicalAttention, 'temporal_attention_block', 'concat/in2');

[Fatih]

thanks alot. It worked, it seems I was just going a bit blind. Now it is solved once I checked it thoroughly.