How to create a custom multi input layer for sequence model

I want to develop similsr type of architechture and Matlab does not support 2 input sequence to one model. Hence could you please let me know some perfect example of deveoping custom layer for inputs like this.

I have also tried to develop the same model in keras and try to import. But it is not helpful as well.

Please help and Thanks in Advance

[Maryam Khairunissa]

hi,

have you solved this issue? I have similar case but for 3 sequence input layers.

[Gawsalyan Sivapalan]

Hi,

Unfortuntely I coudn't get help from MATLAB teams. I tried editing a custom layer but it didn't get accepted by the MLToolbox as it was keep throwing error saying mismatch in size for backpropagation (You can try it as well). Therefore, I developbed my own ML toolbox in matlab for the purpose. https://github.com/gawsalyan/FloatingPoint-NN-from-scratch-MATLAB

Please feel free to contribute to the code base if possible or let me know of any errors as it is solely developed by myself. If you need a fixedpoint version, you can look into my github for the same.

It is very similar to how you create layers in original MATLAB ML Toolbox.

Hope this will help.

P.S.: - Unfortunately I have developed and tested only few layers LSTM and MLP mainly. Only SGD for learning is used.

Cheers!

Siva

Answer 1

[Kenta]

As of 2019b, a new system called "custom training loop" which enables you to implement multi-input CNN is available.

For example, you can refer to the example below.

As of 2019a, to implement multi-input CNN is not easy. If you can update the version of Matlab, please try the example below.

https://jp.mathworks.com/matlabcentral/fileexchange/74760-image-classification-using-cnn-with-multi-input

[Van Vy]

Could you please explain more about the connect layer? If I have 3 inputs, how can I fix the layers? I have trouble with 3 inputs and connectLayers() function. it shows input 1 and input 3 only. Here is my code:

layers2=renameLayer(layers2,'_2');

layersRemoved=[layers(1:end);concatenationLayer(1,3,'Name','cat')];

lgraphAggregated = addLayers(layerGraph(layersRemoved),layers2(1:end));

lgraphAggregated = connectLayers(lgraphAggregated,'fc_2','cat/in2');

layers3=renameLayer(layers3,'_3');

layersRemoved1=[layers2(1:end);concatenationLayer(1,3,'Name','cat')];

lgraphAggregated = addLayers(layerGraph(layersRemoved),layers3(1:end));

lgraphAggregated = connectLayers(lgraphAggregated,'fc_3','cat/in3');