Is selfAttentionLayer in MATLAB2023a only valid for one-dimensional data?

Question

silvia meadows 2023-3-22

2
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1933225-is-selfattentionlayer-in-matlab2023a-only-valid-for-one-dimensional-data

回答： Gayathri 2025-6-13

When I try to combine selfAttentionLayer with ResNet101, it says that 'The input data contains at most one spatial dimension'. It doesn't work even if I put this layer between the 'pool5' and 'fc', which makes the input to selfAttentionLayer 1 times 1 spatially. According to the example on this page, it seems that selfAttentionLayer is only valid for one-dimensional data. What should I do if a want to use it in an image classification task?

THANK YOU so much for your reply! I am new to deep learning and I really need your help!

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

bin 2023-5-13

I'm also wondering about this question, the current version doesn't seem to provide layers to convert images and sequences

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Gayathri 2025-6-13

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1933225-is-selfattentionlayer-in-matlab2023a-only-valid-for-one-dimensional-data#answer_1566479

在 MATLAB Online 中打开

Hi @silvia meadows,

Yes, as you identified "selfAttentionLayer" operates on one dimensional data. So when you have to use the "selfAttentionLayer" for a classification task you can use a "flattenLayer" after a "maxPooling2dLayer" to bring the data to the required format. Please refer to the code below for better understanding. This example uses "DigitDataset" in MATLAB.

% load digit dataset
digitDatasetPath = fullfile(matlabroot, 'toolbox', 'nnet', 'nndemos', 'nndatasets', 'DigitDataset');
imds = imageDatastore(digitDatasetPath, ...
    'IncludeSubfolders', true, 'LabelSource', 'foldernames');
[imdsTrain, imdsValidation] = splitEachLabel(imds, 0.7, 'randomized');
% define network architecture
layers = [
    imageInputLayer([28 28 1], 'Name', 'input')
    convolution2dLayer(3, 32, 'Padding', 'same', 'Name', 'conv1')
    batchNormalizationLayer('Name', 'bn1')
    reluLayer('Name', 'relu1')
    maxPooling2dLayer(2, 'Stride', 2, 'Name', 'maxpool1')
    convolution2dLayer(3, 64, 'Padding', 'same', 'Name', 'conv2')
    batchNormalizationLayer('Name', 'bn2')
    reluLayer('Name', 'relu2')
    maxPooling2dLayer(2, 'Stride', 2, 'Name', 'maxpool2')
    flattenLayer('Name', 'flatten')
    selfAttentionLayer(8, 64, 'Name', 'self_attention')
    fullyConnectedLayer(10, 'Name', 'fc')
    softmaxLayer('Name', 'softmax')
    classificationLayer('Name', 'output')]
% set training options
options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.01, ...
    'MaxEpochs', 5, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', imdsValidation, ...
    'ValidationFrequency', 30, ...
    'Verbose', false, ...
    'Plots', 'training-progress')
% training the network
net = trainNetwork(imdsTrain, layers, options);