Error on Multiple feature Input layers

I created a simple network with 4 featureInputs. The datastore and network were in the following program.

However, when my program call "trainNetwork", I got an error message:

"Input datastore returned more than one observation per row for network input 1."

I know there are few similar questions being asked in the past. But, no one points out exactly where is the problem in the code. Can someone help me take a look and teach me where is the problem?

Thanks.

%%

NumVars = 3;

NumInputs = 4;

TotRows = 2;

%% while hasdata(ds_1), read(ds_1), end

ds_1 = arrayDatastore(zeros(TotRows, NumVars));

ds_2 = arrayDatastore(zeros(TotRows, NumVars - 1));

ds_3 = arrayDatastore(zeros(TotRows, NumVars));

ds_4 = arrayDatastore(zeros(TotRows, NumVars - 1));

TargetY = zeros(TotRows, 1);

ds_TargetY = arrayDatastore(TargetY);

dsTrain = combine(ds_1, ds_2, ds_3, ds_4, ds_TargetY);

%%

lgraph = layerGraph();

for i = 1 : NumInputs

if i == NumInputs

UsedVars = NumVars - 1;

else

UsedVars = NumVars;

end

NameStrIn = ['In_' num2str(i)];

inLayer = featureInputLayer(UsedVars, 'Name', NameStrIn);

lgraph = addLayers(lgraph, inLayer);

NameStrFC = ['FC_' num2str(i)];

fcLayer = fullyConnectedLayer(UsedVars, 'Name', NameStrFC);

lgraph = addLayers(lgraph, fcLayer);

lgraph = connectLayers(lgraph, ['In_' num2str(i)], ['FC_' num2str(i)]);

end

concatLayer = concatenationLayer(1, NumInputs, 'Name', 'Concat');

lgraph = addLayers(lgraph, concatLayer);

for CurIdx = 1 : NumInputs

lgraph = connectLayers(lgraph, ['FC_' num2str(CurIdx)], ['Concat' '/in' num2str(CurIdx)]);

end

Final_Layers = [

fullyConnectedLayer(2, 'Name', 'Final_Layers');

fullyConnectedLayer(1)

regressionLayer];

lgraph = addLayers(lgraph, Final_Layers);

lgraph = connectLayers(lgraph, 'Concat', 'Final_Layers');

analyzeNetwork(lgraph)

%%

options = trainingOptions("sgdm", ...

MaxEpochs=15, ...

InitialLearnRate=0.01, ...

Plots="training-progress", ...

Verbose=0);

net = trainNetwork(dsTrain, lgraph, options);

[Chih]

Slightly updated program:

NumVars = 3;

NumInputs = 4;

TotRows = 2;

%% while hasdata(ds_1), read(ds_1), end

ds_1 = arrayDatastore(zeros(TotRows, NumVars));

ds_2 = arrayDatastore(zeros(TotRows, NumVars));

ds_3 = arrayDatastore(zeros(TotRows, NumVars));

ds_4 = arrayDatastore(zeros(TotRows, NumVars - 1));

TargetY = zeros(TotRows, 1);

ds_TargetY = arrayDatastore(TargetY);

dsTrain = combine(ds_1, ds_2, ds_3, ds_4, ds_TargetY);

%%

lgraph = layerGraph();

for i = 1 : NumInputs

if i == NumInputs

UsedVars = NumVars - 1;

else

UsedVars = NumVars;

end

NameStrIn = ['In_' num2str(i)];

inLayer = featureInputLayer(UsedVars, 'Name', NameStrIn);

lgraph = addLayers(lgraph, inLayer);

NameStrFC = ['FC_' num2str(i)];

fcLayer = fullyConnectedLayer(UsedVars, 'Name', NameStrFC);

lgraph = addLayers(lgraph, fcLayer);

lgraph = connectLayers(lgraph, ['In_' num2str(i)], ['FC_' num2str(i)]);

end

concatLayer = concatenationLayer(1, NumInputs, 'Name', 'Concat');

lgraph = addLayers(lgraph, concatLayer);

for CurIdx = 1 : NumInputs

lgraph = connectLayers(lgraph, ['FC_' num2str(CurIdx)], ['Concat' '/in' num2str(CurIdx)]);

end

Final_Layers = [

fullyConnectedLayer(2, 'Name', 'Final_Layers');

fullyConnectedLayer(1)

regressionLayer];

lgraph = addLayers(lgraph, Final_Layers);

lgraph = connectLayers(lgraph, 'Concat', 'Final_Layers');

analyzeNetwork(lgraph)

%%

options = trainingOptions("sgdm", ...

MaxEpochs=15, ...

InitialLearnRate=0.01, ...

Plots="training-progress", ...

Verbose=0);

net = trainNetwork(dsTrain, lgraph, options);

Answer 1

[Vinayak Choyyan]

Hello Chih,

As per my understanding, you have designed a neural network with 4 ‘featureInputLayer’. As dummy data, you are providing the model with 3 ‘arrayDatastore’of size 2x3 and a 4th ‘arrayDatastore’ of size 2x2. TargetY is the output of size 2x1. I assume these 2 rows are separate observations and hence making the input to the model of size 11 and output of size 1. To feed the training data to the model, you have used ‘combine’ to create a ‘CombinedDatatore’.

When you try to train a network using ‘trainNetwork’, it reads data from the ‘CombinedDatatore’. If we check what is being returned from the ‘CombinedDatatore’, we can see that it returns the following.

NumVars = 3;
NumInputs = 4;
TotRows = 2;

ds_1 = arrayDatastore(zeros(TotRows, NumVars));
ds_2 = arrayDatastore(zeros(TotRows, NumVars));
ds_3 = arrayDatastore(zeros(TotRows, NumVars));
ds_4 = arrayDatastore(zeros(TotRows, NumVars - 1));
TargetY = zeros(TotRows, 1);
ds_TargetY = arrayDatastore(TargetY);
dsTrain = combine(ds_1, ds_2, ds_3, ds_4, ds_TargetY);
read(dsTrain)

ans = 1×5 cell array
    {[0 0 0]}    {[0 0 0]}    {[0 0 0]}    {[0 0]}    {[0]}

There are 4 inputs and 1 output to the model. Hence ‘trainNetwork’ is expecting to get 4+1 = 5 data points. You can see that the first 3 elements are cells of size 1x3. But the model is expecting 3x1.

As a workaround for this issue you faced, please try the below code. I have only modified the 5 ‘arrayDatastore’ you created to return the expected input size data. The rest is the same.

clc;clear;

NumVars = 3;
NumInputs = 4;
TotRows = 2;
%% while hasdata(ds_1), read(ds_1), end
tmp=zeros(TotRows, NumVars);
tmp1=zeros(TotRows, NumVars-1);
ds_1 = arrayDatastore(tmp',"IterationDimension",2);
ds_2 = arrayDatastore(tmp',"IterationDimension",2);
ds_3 = arrayDatastore(tmp',"IterationDimension",2);
ds_4 = arrayDatastore(tmp1',"IterationDimension",2);
TargetY = zeros(TotRows, 1);
ds_TargetY = arrayDatastore(TargetY',"IterationDimension",2);
dsTrain = combine(ds_1, ds_2, ds_3, ds_4, ds_TargetY);
%%
lgraph = layerGraph();
for i = 1 : NumInputs
    if i == NumInputs
        UsedVars = NumVars - 1;
    else
        UsedVars = NumVars;
    end
    NameStrIn = ['In_' num2str(i)];
    inLayer = featureInputLayer(UsedVars, 'Name', NameStrIn);
    lgraph = addLayers(lgraph, inLayer);
    NameStrFC = ['FC_' num2str(i)];
    fcLayer = fullyConnectedLayer(UsedVars, 'Name', NameStrFC);
    lgraph = addLayers(lgraph, fcLayer);
    lgraph = connectLayers(lgraph,  ['In_' num2str(i)], ['FC_' num2str(i)]);
end
concatLayer = concatenationLayer(1, NumInputs, 'Name', 'Concat');
lgraph = addLayers(lgraph, concatLayer);
for CurIdx = 1 : NumInputs 
    lgraph = connectLayers(lgraph, ['FC_' num2str(CurIdx)], ['Concat' '/in' num2str(CurIdx)]);
end
Final_Layers = [
    fullyConnectedLayer(2, 'Name', 'Final_Layers');    
    fullyConnectedLayer(1)
    regressionLayer];
lgraph = addLayers(lgraph, Final_Layers);
lgraph = connectLayers(lgraph, 'Concat', 'Final_Layers');
analyzeNetwork(lgraph)
%%
options = trainingOptions("sgdm", ...
    MaxEpochs=15, ...
    InitialLearnRate=0.01, ...
    Plots="training-progress", ...
    Verbose=0);
net = trainNetwork(dsTrain, lgraph, options);

If you would like to read more on the following, please check out these documentation pages:

• CombinedDatastore Datastore to combine data read from multiple underlying datastores - MATLAB - MathWorks India
• trainNetwork Train deep learning neural network - MATLAB trainNetwork - MathWorks India
• ArrayDatastore Datastore for in-memory data - MATLAB - MathWorks India

I hope this resolves the issue you are facing.