3D convolutional auto encoder

4 次查看（过去 30 天）

Juuso Korhonen 2021-3-29

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/786496-3d-convolutional-auto-encoder

I try to define a 3D convolutional auto encoder, but when trying to train my training error just stays the same. Here is my structure:

layers = [
    image3dInputLayer([120 160 120 1], 'Name', 'Input', 'Normalization',"none") %??
    convolution3dLayer([3 3 3], 10, 'Name', 'conv1', "Padding", "same", 'Stride', [1 1 1])
    reluLayer('Name', 'relu1')
    maxPooling3dLayer([2 2 2], "Stride", [2 2 2], "Name", 'max1')
    convolution3dLayer([3 3 3], 10, 'Name', 'conv2', "Padding", "same", 'Stride', [1 1 1])
    reluLayer('Name', 'relu2')
    maxPooling3dLayer([2 2 2], "Stride", [2 2 2], "Name", 'max2')
    convolution3dLayer([3 3 3], 10, 'Name', 'conv3', "Padding", "same", 'Stride', [1 1 1])
    reluLayer('Name', 'relu3')
    maxPooling3dLayer([2 2 2], "Stride", [2 2 2], "Name", 'max3')
    dropoutLayer(0.5, "Name", 'drop')
    convolution3dLayer([3 3 3], 10, 'Name', 'conv4', "Padding", "same", 'Stride', [1 1 1])
    reluLayer('Name', 'relu4')
    resize3dLayer("OutputSize",[30 40 30], 'Method', "nearest", 'Name','up1')
    convolution3dLayer([3 3 3], 10, 'Name', 'conv5', "Padding", "same", 'Stride', [1 1 1])
    reluLayer('Name', 'relu5')
    resize3dLayer("OutputSize",[60 80 60], 'Method',"nearest", 'Name','up2')
    convolution3dLayer([3 3 3], 10, 'Name', 'conv6', "Padding", "same", 'Stride', [1 1 1])
    reluLayer('Name', 'relu6')
    convolution3dLayer([1 1 1], 1, 'Name', 'concatenation', "Padding", "same", 'Stride', [1 1 1])
    resize3dLayer("OutputSize",[120 160 120], 'Method', "nearest", 'Name','up31') %??
    sigmoidLayer('Name', 'sigmoid') %??
    regressionLayer('Name', 'reglayer')
    ];
lgraph = layerGraph(layers);

Especially I would like to know if the combining of the final feature maps is done correctly with the last convolution3dLayer, since I couldn't find a suitable layer for it in 3D. Also, if the use of sigmoid before the output is suitable? If I use normal reluLayer, the training error just goes to Inf.

I do the training the following way:

imdsTrain = transform(imdsTrain, @resize2AE,'IncludeInfo',true);
[net, info] = trainNetwork(imdsTrain, lgraph, options);
% Where the resize2AE function is 
function [dataOut,info] = resize2AE(data,info)
    numRows = size(data,1);
    dataOut = cell(numRows,2);
    % since ReadSize is expected to be >1, data comes in cell form containing
    % multiple images
    for idx = 1:numRows
        % get the image out of the datacell
        D = data{idx,1}; 
        D = single(D);
        D = imresize3(D, [120, 160, 120]);
        D = normalize(D, 'range'); % normalize to [0,1] range
        dataOut(idx,:) = {D,D}; % This is supposed to work for AE!!!!!!!
    end
end