Training Data type error for a CNN using trainnet function

Trying to use a convolution1dLayer for my sequence input data put when I try to train it i get the error:

"Error using trainnet

Invalid targets. Network expects numeric or categorical targets, but received a cell array."

I've looked at many exemples of how the data must be structed but even if is in the same format, it doesn't work.

For the predictors I'm doing a test with only 4 observations, each one with 4 features and 36191 points:

For the targets there are also for observations with only one target each and also 36191 points:

I can't understand why it doesn't accept it, like I said, its equal to many other exemples. I leave down here the code for the CNN-LSTM network and the trainnet function:

lgraph = layerGraph();
tempLayers = [
    sequenceInputLayer(4,"Name","input")
    convolution1dLayer(4,32,"Name","conv1d","Padding","same")
    globalAveragePooling1dLayer("Name","gapool1d")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = lstmLayer(25,"Name","lstm");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = lstmLayer(25,"Name","lstm_1");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    concatenationLayer(1,2,"Name","concat")
    lstmLayer(55,"Name","lstm_2")
    dropoutLayer(0.5,"Name","drop")
    fullyConnectedLayer(1,"Name","fc")
    sigmoidLayer("Name","sigmoid")];
lgraph = addLayers(lgraph,tempLayers);

% clean up helper variable
clear tempLayers;

lgraph = connectLayers(lgraph,"gapool1d","lstm");
lgraph = connectLayers(lgraph,"gapool1d","lstm_1");
lgraph = connectLayers(lgraph,"lstm","concat/in1");
lgraph = connectLayers(lgraph,"lstm_1","concat/in2");

plot(lgraph);

epochs = 800;
miniBatchSize = 128;
LRDropPeriod = 200; 
InitialLR = 0.01;
LRDropFactor = 0.1; 
valFrequency = 30; 

options = trainingOptions("adam", ...
    MaxEpochs=epochs, ...
    SequencePaddingDirection="left", ...
    Shuffle="every-epoch", ...
    GradientThreshold=1, ...
    InitialLearnRate=InitialLR, ...
    LearnRateSchedule="piecewise", ...
    LearnRateDropPeriod=LRDropPeriod, ...
    LearnRateDropFactor=LRDropFactor, ...
    MiniBatchSize=miniBatchSize, ...
    Plots="training-progress", ...
    Metrics="rmse", ...
    Verbose=0, ...
    ExecutionEnvironment="parallel");


CNN_LTSM = trainnet(trainDataX, trainDataY, dlnetwork(lgraph),"mse",options);

using version 2023b

Answer 1

[Milan Bansal]

Hi Zowie Silva,

I understand that you are facing an error while using trainnet function for training a sequence-to-sequence neural network problem.

Use "analyzeNetwork" to look into the structure of your neural network.

From the above screenshot, it can be observed that the model is losing the "T" dimension after "gapool1D" layer. Due to which the model is expecting a numerical or categorical target value instead of sequence.

The Neural Network built by you is using globalAveragePooling1dLayer. As per the documentation of globalAveragePooling1dLayer, for time series and vector sequence input (data with three dimensions corresponding to the "C" (channel), "B" (batch), and "T" (time) dimensions), the layer pools over the "T" (time) dimension. That means, after the pooling, only "C" and "B" dimension will remain.

If you wish to do the average pooling along channel dimension, you can create a functionLayer and replace the "gapool1D" layer with it. Please refer to the following code snippet to modify your model:

lgraph = layerGraph();

functionLayer = functionLayer(@(X) mean(X, 1), 'Name', 'avgPoolChannel'); % defining function layer

tempLayers = [
    sequenceInputLayer(4,"Name","input")
    convolution1dLayer(4,32,"Name","conv1d","Padding","same") % using the function layer
    functionLayer];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = lstmLayer(25,"Name","lstm");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = lstmLayer(25,"Name","lstm_1");
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    concatenationLayer(1,2,"Name","concat")
    lstmLayer(55,"Name","lstm_2")
    dropoutLayer(0.5,"Name","drop")
    fullyConnectedLayer(1,"Name","fc")
    sigmoidLayer("Name","sigmoid")];
lgraph = addLayers(lgraph,tempLayers);

% clean up helper variable
clear tempLayers;

lgraph = connectLayers(lgraph,"avgPoolChannel","lstm"); % connect to function layer
lgraph = connectLayers(lgraph,"avgPoolChannel","lstm_1"); % connect to function layer

lgraph = connectLayers(lgraph,"lstm","concat/in1");
lgraph = connectLayers(lgraph,"lstm_1","concat/in2");

Please refer to the following documentation links to learn more about:

• globalAveragePooling1dLayer: https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/nnet.cnn.layer.globalaveragepooling1dlayer.html
• functionLayer: https://www.mathworks.com/help/releases/R2023b/deeplearning/ref/nnet.cnn.layer.functionlayer.html
• trainnet: https://in.mathworks.com/help/releases/R2023b/deeplearning/ref/trainnet.html

Hope this helps!

[Zowie Silva]

Hi Milan.

Firstly, thank you for your support and explanation!

I've tried your code and I've received the following error:

"Error using deep.internal.train.Trainer/train

Layer 'input': Invalid input data. Invalid size of channel dimension. Layer expects input with channel dimension size 4 but received input with size 36191.

Error in deep.internal.train.ParallelTrainer>iTrainWithSplitCommunicator (line 158)

remoteNetwork = train(remoteTrainer, remoteNetwork, workerMbq);

Error in deep.internal.train.ParallelTrainer/computeTraining (line 119)

spmd

Error in deep.internal.train.Trainer/train (line 33)

[net, trainerInfo] = computeTraining(trainer, net, mbq);

Error in deep.internal.train.trainnet (line 51)

net = train(trainer, net, mbq);

Error in trainnet (line 91)

[net,info] = deep.internal.train.trainnet(mbq, net, loss, options, ..."

But to solve it I've just change de dimensions of the cells in TrainDataX and TrainDataY to (36191x4 and 36191x1, respectively)

Again, thanks a lot for the help, this problem has been quite the headache.