How to design a locally connected layer for use in a convolution neural network??

I am trying to replicate the architecture found in this paper (https://www.nature.com/articles/srep36571), but the deepNetworkDesigner app does not have a locally connected layer. The best alternative I found is to use the 2d convolution but with 1x1 filter size to approximate the 1d convolution behavior (https://stackoverflow.com/questions/50388014/1d-convolution-for-cnn). According to (https://keras.io/api/layers/locally_connected_layers/locall_connected1d/) the locally connected layer is similar to 1d convolution, except the weights are unshared.

How can I go about actually making a proper 1d convolution layer?

Answer 1

[Srivardhan Gadila]

You can refer to Define Custom Deep Learning Layers & Deep Learning Custom Layers and implement your own custom deep learning layer.

[Michael Houston]

Thank you, Srivardhan. While I will definitely be using this to create the custom layer my main issue is the construction of the layer itself. I am not too familiar with deep learning but I appreciate any tips/hints to help me design the locally connected layer!