Feed inputs to inner layers of a neural network (custom structure)

Question

Deskj 2022-5-11

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1716840-feed-inputs-to-inner-layers-of-a-neural-network-custom-structure

回答： T.Nikhil kumar 2023-9-29

I was wondering if Matlab allows to feed the inputs of a neural network to inner layers in a custom structure (meaning not only to the first input layer).

Here is an example of what I mean:

In this example two inputs are fed to the first block (input layer x1-x2, hidden layer, output layer u1);

Then the second block is fed with the output of the first block and other 2 inputs (input layer u1-x3-x4, hidden layer, output layer y);

Then the final output is predicted.

Is there a way to create this kind of structure with a single model?

I tried to create it using the dlnetworks command, but it obviously didn't work because multiple input layers are not supported.

(And I wouldn't even know how the training command should be written at this point)

net = dlnetwork(layers);
layers = [
    %First block
    sequenceInputLayer(2,"Name","Block1") %x1, x2
    fullyConnectedLayer(3,"Name","fc_11")
    fullyConnectedLayer(1,"Name","fc_12")
    regressionLayer("Name","u1") %u1
    
    %Second block
    sequenceInputLayer(3,"Name","Block2") %u1, x3, x4
    fullyConnectedLayer(3,"Name","fc_21")
    fullyConnectedLayer(1,"Name","fc_22")
    regressionLayer("Name","Y") %y
    ];

There is a lot of documentation about custom structures and custom training loops, but I feel kind of lost and I do not know where I should have a look at.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

T.Nikhil kumar 2023-9-29

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1716840-feed-inputs-to-inner-layers-of-a-neural-network-custom-structure#answer_1321919

Hello,

I understand that you want to feed inputs to a custom network at different levels i.e. at the inner layers of the network.

It is possible to create networks that can take multiple inputs and process them at different stages or layers of the network. This can be done by creating separate branches for each set of inputs and then connecting them to form the final network. But note that a network can have at most one sequence input layer. Hence, you cannot create the structure you wish with 2 sequence input layers.

You can try using other types of Input Layers for this purpose.

Please find below the procedure to create a network with multiple image inputs at inner layers using “imageInputLayer”:

Define a set of layers with the first set of inputs and create a network using the “layerGraph” function.
Create a set of layers for the next set of inputs and then add these layers to the first branch’s “layerGraph” using the “addLayers” function.
Connect Layers of the network using “connectLayers” function.