How can a 2D convolution transpose (deconvolution) layer be defined using Neural Network Toolbox?

Dear Matlab community,

It is easy to define Neural Network layers in Matlab for 2D convolutional layers by using the command

convlayer = convolution2dLayer(filterSize,numFilters)

Actually, I want to create a layer that does the opposite: A convolutional transpose or also called deconvolutional layer that maps from the latent space to the image space. A NN consisting of such a layer can be used to generate images. I'm quite new to that topic and I was not very successful using third party libraries like MatConvNet yet. Is it possible to write such a deconvolutional layer that can also interact with the other functions provided by the Neural Network Toolbox, and if yes how could this be done?

Thank you very much in advance! Sebastian

Answer 1

[Adan Garcia]

https://la.mathworks.com/help/nnet/ref/nnet.cnn.layer.transposedconv2dlayer.html