Undefined function handle error when loading a neural network that contain a function layer

Hi everyone,

I created a neural network that uses a functionLayer that I defined by :

functionLayer(@Feature_wise_LM,Name="FiLM_64",Formattable=1,NumInputs=2,NumOutputs=1)

However, when I save the trained neural network with :

save("./trained_networks/FiLM", "trained_network");

and reload it in the same script with :

trained_network = load("./trained_networks/FiLM.mat");
trained_network = trained_network.trained_network;

I get the error :

Warning: While loading an object of class 'dlnetwork':
Error using nnet.internal.cnn.layer.GraphExecutor/propagate (line 354)
Execution failed during layer(s) 'FiLM_64'.

Error in deep.internal.network.ExecutableNetwork/configureForInputsAndForwardOnLayer (line 347)
propagate(this, fcn, Xs, outputLayerIdx, outputLayerPortIdx);

Error in deep.internal.network.EditableNetwork/convertToDlnetwork (line 101)
[executableNetwork, layerOutputSizes] = configureForInputsAndForwardOnLayer(...
    
Error in dlnetwork.loadobj (line 741)
net = convertToDlnetwork(privateNet, exampleInputs, initializeNetworkWeights);

Caused by:
Undefined function handle.

Error in nnet.cnn.layer.FunctionLayer/predict (line 61)
[varargout{1:layer.NumOutputs}] = layer.PredictFcn(varargin{:}); 

I tried to save the function in a dedicated file "Feature_wise_LM.m" , but it didn't work

Answer 1

[Matt J]

I tried to save the function in a dedicated file "Feature_wise_LM.m" , but it didn't work

It should work. Make sure the location of that file is on the Matlab path.

Make sure as well that it is in the same location as when the functionLayer was created, though I'm not sure it is absolutely necessary.

[Julien]

Hello Matt,

Thank you for your answer!

I am quite sure that it doesn't work like this. The only workaround I found is that if I put

functionLayer(@(X1,X2) Feature_wise_LM(X1,X2),Name="FiLM_64",Formattable=1,NumInputs=2,NumOutputs=1)

it for some reason works.

However, a complete training of the model takes a very very long time, so I wonder if there is a way to recover the saved model that I trained by declaring the layer as

functionLayer(@Feature_wise_LM,Name="FiLM_64",Formattable=1,NumInputs=2,NumOutputs=1)

[Matt J]

The problem has to be something in your environment. Here it works fine:

net=dlnetwork(functionLayer((@testFun),Name="FiLM_64",Formattable=1,NumInputs=2,NumOutputs=1), ...
             dlarray(1,'SSC'),dlarray(2,'SSC'));

save('netfile.mat', 'net')

clear

load('netfile.mat', 'net')

net,

net = 
  dlnetwork with properties:

         Layers: [1×1 nnet.cnn.layer.FunctionLayer]
    Connections: [0×2 table]
     Learnables: [0×3 table]
          State: [0×3 table]
     InputNames: {'FiLM_64/in1'  'FiLM_64/in2'}
    OutputNames: {'FiLM_64'}
    Initialized: 1

  View summary with summary.

function Y=testFun(X1,X2)

  Y=X1+X2;

end

[Julien]

Hello Matt,

The code indeed works on matlab online.

I tested it on my personal computer, it worked well, and it even worked when I replaced

clear

by

clear all

However, if I turn matlab off, and reopen it, I have the usual error when executing

load('netfile.mat', 'net')

[Matt J]

If so, it means your Matlab path changes when you close and reopen Matlab.

What happens, after reopening Matlab, when you do,

which Feature_wise_LM

from the command line?

[Julien]

It does give me the right path...

However, in your example the function is in the same script, and when defining the layer by "@(X1,X2) Feature_wise_LM(X1,X2)", it works even when turning matlab off and reopening it, so I think that matlab can access the function

[Matt J]

However, in your example the function is in the same script.

The example didn't require that. Here it is again, with testFun defined in its own mfile, as you claim you are doing. It still works, even without wrapping the handle in an anonymous function:

net=dlnetwork(functionLayer(@testFun,Name="FiLM_64",Formattable=1,NumInputs=2,NumOutputs=1), ...
             dlarray(1,'SSC'),dlarray(2,'SSC'));

save('netfile.mat', 'net')

clear

load('netfile.mat', 'net')

net,

net = 
  dlnetwork with properties:

         Layers: [1×1 nnet.cnn.layer.FunctionLayer]
    Connections: [0×2 table]
     Learnables: [0×3 table]
          State: [0×3 table]
     InputNames: {'FiLM_64/in1'  'FiLM_64/in2'}
    OutputNames: {'FiLM_64'}
    Initialized: 1

  View summary with summary.