How to train a vector multiple-input multiple-output network

I want to train a vector multi-input, multi-output network, but I get the error “Number of input data formats (1) and number of network inputs (2) must match”. This is the code,

clear
input1ds=signalDatastore("input1.csv")
input2ds=signalDatastore("input2.csv")
output1ds=signalDatastore("output1.csv")
output2ds=signalDatastore("output2.csv")
ds=combine(input1ds,input2ds,output1ds,output2ds)
isShuffleable(ds)

% 入力とターゲット
% x = [[0:0.1:10]' [0:0.1:10]']
% t = sin(x)

net=dlnetwork

% ニューラルネットワークの定義
layers1 = [
    featureInputLayer(1,"name","input1")
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(1) % 1ユニットの全結合層
];
layers2 = [
    featureInputLayer(1,"Name","input2")
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(1) % 1ユニットの全結合層
];

net=addLayers(net,layers1)
net=addLayers(net,layers2)
net.plot
% layers=layerGraph()
% layers=addLayers(layers,layers1)
% layers=addLayers(layers,layers2)




% オプションの設定
options = trainingOptions('sgdm', ... % 最適化アルゴリズム
    'MaxEpochs', 500, ... % 最大エポック数
    'MiniBatchSize', 2^3, ... % ミニバッチサイズ
    'Verbose', true,...% 進行状況の表示
    'InputDataFormats', {'SB'} ...
    )

% ニューラルネットワークのトレーニング
customLossFunction = @(Y, T) customloss(Y, T);
net = trainnet(ds,net,customLossFunction,options) % x,tは縦ベクトル

The “input1.csv” and “output1.csv” all contain a single column of vertical vectors. This time, it is a 2-input, 2-output network.

The function “customloss” is a function I defined myself. This is the error statement.

Error using trainnet (line 46)

Number of input data formats (1) and number of network inputs (2) must match.

Error in parallel_learn_test (line 53)

net = trainnet(ds,net,customLossFunction,options) % x,tは縦ベクトル

What is wrong? And what solutions are available?

Answer 1

[Matt J]

You have only entered one InputDataFormat. The error message is telling you that you need two of them, e.g.,

options = trainingOptions(______
    'InputDataFormats', {'SB','SB'} ...
    )

because the network has two inputs.

[Kennosuke Nakabayashi]

Thank you for answering my question. You have helped me resolve this error. However, I got a new error statement. The change is that I defined the datastore with the arrayDatastore function.

% 入力とターゲット
x = [[0:0.1:10]' [0:0.1:10]']
t = sin(x)
ds=combine(arrayDatastore(x(:,1),"ReadSize",length(x(:,1))),...
    arrayDatastore(x(:,2),"ReadSize",length(x(:,2))),...
    arrayDatastore(t(:,1),"ReadSize",length(t(:,1))),...
    arrayDatastore(t(:,2),"ReadSize",length(t(:,2))))

net=dlnetwork

% ニューラルネットワークの定義
layers1 = [
    featureInputLayer(1,"name","input1")
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(1) % 1ユニットの全結合層
];
layers2 = [
    featureInputLayer(1,"Name","input2")
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(1) % 1ユニットの全結合層
];

net=addLayers(net,layers1)
net=addLayers(net,layers2)
net.plot
% layers=layerGraph()
% layers=addLayers(layers,layers1)
% layers=addLayers(layers,layers2)




% オプションの設定
options = trainingOptions('sgdm', ... % 最適化アルゴリズム
    'MaxEpochs', 500, ... % 最大エポック数
    'MiniBatchSize', 2^3, ... % ミニバッチサイズ
    'Verbose', true,...% 進行状況の表示
    'InputDataFormats', {'SB','SB'} ...
    )

% ニューラルネットワークのトレーニング
customLossFunction = @(Y, T) customloss(Y, T);
net = trainnet(ds,net,customLossFunction,options) % x,tは縦ベクトル

The statement of that error is this.

Error using trainnet (line 46)

Layer 'input1': Invalid input data. Invalid number of spatial dimensions. Layer expects 0 but received 1.

Error in parallel_learn_test (line 57)

net = trainnet(ds,net,customLossFunction,options) % x,tは縦ベクトル

Sorry for asking so many questions.

[Matt J]

Use 'CB' as the input format ReadSize=1,

% 入力とターゲット
x = [[0:0.1:10]' [0:0.1:10]'];
t = sin(x);
N=height(x);
fcn=@(z)  arrayDatastore(z,'ReadSize',1);
ds=combine( fcn(x(:,1)), fcn(x(:,2))  , fcn(t(:,1)), fcn(t(:,2))   );

p=ds.preview

p = 1x4 cell array
    {[0]}    {[0]}    {[0]}    {[0]}

net=dlnetwork;

% ニューラルネットワークの定義
layers1 = [
    featureInputLayer(1,"name","input1")
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(1) % 1ユニットの全結合層
];
layers2 = [
    featureInputLayer(1,"Name","input2")
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(10) % 10ユニットの全結合層
    tanhLayer
    fullyConnectedLayer(1) % 1ユニットの全結合層
];

net=addLayers(net,layers1);
net=addLayers(net,layers2);
% layers=layerGraph()
% layers=addLayers(layers,layers1)
% layers=addLayers(layers,layers2)




% オプションの設定
options = trainingOptions('sgdm', ... % 最適化アルゴリズム
    'MaxEpochs', 10, ... % 最大エポック数
    'MiniBatchSize', 2^3, ... % ミニバッチサイズ
    'Verbose', true,...% 進行状況の表示
    'InputDataFormats', {'CB','CB'} ...
    );

% ニューラルネットワークのトレーニング
customLossFunction = @(Y1,Y2, T1,T2) sum(abs([Y1,Y2]-[T1,T2]),'all');
net = trainnet(ds,net,customLossFunction,options) % x,tは縦ベクトル

    Iteration    Epoch    TimeElapsed    LearnRate    TrainingLoss
    _________    _____    ___________    _________    ____________
            1        1       00:00:01         0.01          9.3311
           50        5       00:00:02         0.01          9.9611
          100        9       00:00:03         0.01          8.9918
          120       10       00:00:03         0.01          8.9776
Training stopped: Max epochs completed

net = 
  dlnetwork with properties:

         Layers: [12x1 nnet.cnn.layer.Layer]
    Connections: [10x2 table]
     Learnables: [12x3 table]
          State: [0x3 table]
     InputNames: {'input1'  'input2'}
    OutputNames: {'fc_3'  'fc_6'}
    Initialized: 1

  View summary with summary.