custom neural network error( About dlfeval )

I am running a custom neural network using a low level api ( dlnetwork, dlfeval, adamupdate, dlgradient).

However, while running dlfeval, I get an error.

error is here.

Error using dlfeval (line 43)

First input argument must be a formatted dlarray.

Error in deep (line 639)

[gradient,loss]=dlfeval(@modelGradients,dlnet,dlX);

I think it's an error about dlarray, but if you look at my code, I declare dlarray in the input.

my code is here. (Input feature:8 // Target:1)

clear,clc,close all

data=readmatrix('train.csv');

inputs=data(:,1:8);

targets=data(:,9);

input2=transpose(inputs);

target2=transpose(targets);

inputs2=normalize(input2,2,'range');

layers= [sequenceInputLayer([8],'Name','input')

fullyConnectedLayer(64,'Name','fc1')

tanhLayer('Name','tanh1')

fullyConnectedLayer(32,'Name','fc2')

tanhLayer('Name','tanh2')

fullyConnectedLayer(16,'Name','fc3')

tanhLayer('Name','tanh3')

fullyConnectedLayer(8,'Name','fc4')

tanhLayer('Name','tanh4')

fullyConnectedLayer(1,'Name','fc5')

];

lgraph=layerGraph(layers);

dlnet=dlnetwork(lgraph);

for it=1:5000

dlX=dlarray(inputs2);

[gradient,loss]=dlfeval(@modelGradients,dlnet,dlX);

dlnet=adamupdate(dlnet,gradient);

end

function [gradient,loss]=modelGradients(dlnet,dlx,t)

out=forward(dlnet,dlx);

loss=immse(out,t);

loss=dlarray(loss);

gradient=dlgradient(loss,dlnet.Learnables);

end

Thanks for reading my question!

Answer 1

[Srivardhan Gadila]

The input dlX for the forward(dlnet,dlX) function should be a formatted dlarray. Refer to the documentation of forward (specifically dlX under Input Arguments) for more information.

Also from the above code, the modelGradients takes in dlnet, dlx and t as input arguments

function [gradient,loss]=modelGradients(dlnet,dlx,t)

but in the for loop to compute the gradients you are not providing the target data i.e., target2 as dlarray.

[gradient,loss]=dlfeval(@modelGradients,dlnet,dlX);

due to which you may get another error w.r.t

loss=immse(out,t);

Refer to the example Train Network Using Custom Training Loop for more information.

[jaehong kim]

Thank you for your answer!

Can you give me an answer for multiple output cases?

For example, the case where the outputsize of the last fullyconnected layer is 6.

fullyConnectedLayer(1,'Name','fc5') ----> fullyConnectedLayer(6,'Name','fc5')