Architecture of convolutional autoencoders in Matlab 2019b

Hello all,

I am very interested in training convolutional autoencoders in MATLAB 2019b.

I have found the instruction trainAutoencoder, but it does not allow to specify the convolutional layers architecture.

I want to design my autoencoder using Deep Network Designer tool, and then train it just as it is done with CNNs, FasterRCNN algorithms, etc, using my image dataset.

Is this possible? Is there any code oxample out there to do so?

Thank you all in advance,

Best,

Alberto

Answer 1

[Mahesh Taparia]

Hi

You can define custom architecture of auoencoder using deep learning layers. You can refer to this documentation for the list of deep learning layers supported in MATLAB. For example, the autoencoder network can be defined as:

layers=[
    imageInputLayer(size,"Name","imageinput",'Normalization','none')  %size is the size of input
    fullyConnectedLayer(9*R,"Name","fc_1")    %R can be any number/ factor
    leakyReluLayer(0.01,"Name","leakyrelu_1")
    fullyConnectedLayer(6*R,"Name","fc_2")
    leakyReluLayer(0.01,"Name","leakyrelu_3")
    fullyConnectedLayer(3*R,"Name","fc_3")
    leakyReluLayer(0.01,"Name","leakyrelu_4")
    fullyConnectedLayer(R,"Name","fc_4")
    batchNormalizationLayer("Name","batchnorm")
    reluLayer('Name','relu1')
    dropoutLayer('Name','drop')
    fullyConnectedLayer(3*R,"Name","fc_4")
    leakyReluLayer(0.01,"Name","leakyrelu_5")
    fullyConnectedLayer(6*R,"Name","fc_4")
    leakyReluLayer(0.01,"Name","leakyrelu_6")
    fullyConnectedLayer(9*R,"Name","fc_4")
    leakyReluLayer(0.01,"Name","leakyrelu_7")
    fullyConnectedLayer(size,"Name","fc_5")
    leakyReluLayer(0.01,"Name","leakyrelu_8")
    regressionLayer("Name","regressionoutput")];

You can use 2D / 3D conv layer/ any other layer as per your architecture. After defining the network, you can train the model. For reference you can check this documentation for that.

For custom loss function, you can check this documentation.

Hope it will help!

[Kenneth Afebu]

@Mahesh Taparia.

Thanks for your previous contribution.

I am new to convolutional autoencoders but it seems to me that the array of layers you gave above do not actually represent a convolutional autoencoder. There are no convolutional layers and I can not pick the encoding and decoding parts of the layers.

Please can you help with a typical convolutional autoencoder layers array designed for MATLAB while also pointing out the encoding and decoding part.

I have been trying to do so but I have not been able as I am new to it.

Thanks in advance.

[Mahesh Taparia]

Hi

You can replace the fully connected layer with convolution layer in the above code, then it will be a convolutional autoencoder. You can refer to this documentation of convolution layer for more details. You can change the layers as per your application.

[Jian Wang]

Hi, Mahesh Taparia. Thank you for your answer. I want to know how can I train the autoencoder in Matlab? I use the following code:

net = trainNetwork(X, X, AE_Layers, AE_options);

X is a array with 13457-by-7, i.e. 7 variables.

But it seems wrong when I get the reconstructed X by:

XReconstructed = predict(net, X);

XReconstructed is full of NANs. Whether the code is wrong? Thank you.