How to predict with 1Dconvolution layers: get 3 channels to produce one output?

Hello there I am tying to build a model to test what a 1D convolution layer will do to my data. I have 3 input accelerometer mesurements in the x,y,z directions and am trying to predict a continuous output. The Xtrain is a 3x500 double with the 3 input accelerometer mesurements over 500 time steps and the Ytrain is a 1x500 double of the target output variable. The code I have developed is here:

numFilters = 64;
filterSize = 5;
droupoutFactor = 0.005;
numBlocks = 4;
dilationFactor = 1;

net = dlnetwork;

layers = sequenceInputLayer(NumFeatures,Normalization="rescale-symmetric",Name="inputaccelerometers")
convolution1dLayer(filterSize,numFilters,DilationFactor=dilationFactor,Padding="causal",Name="conv1_")

net = addLayers(net,layers)
    
options = trainingOptions("adam", ...
    MaxEpochs=60, ...
    miniBatchSize=1, ...
    InputDataFormats="CTB", ...
    Plots="training-progress", ...
    Metrics="rmse", ...
    Verbose=0);


net = trainnet(Xtrain',Ytrain',net,"mse",options)

But when I try and train the model I get and error that says: Number of channels in predictions (3) must match the number of channels in the targets (1). Do I need to add another layer to convert the inputs into one predicted output and which layer should I use to do this? Any suggestions on how to remove this error and get the network to run accurately? Thanks so much!

Answer 1

[Milan Bansal]

Hi Isabelle Museck,

The error you're encountering is due to a mismatch between the number of channels in your predictions and the number of channels in your targets. Since your target is a single continuous output (a 1x500 double), you need to ensure that your network's final output layer produces a single output channel.

To resolve this, you can add a fully connected layer with a single output neuron after your convolutional layers. This fully connected layer will map the multi-channel output of the convolutional layers to a single continuous output.

Here's a revised version of your code to include this adjustment:

numFilters = 64;
filterSize = 5;
dropoutFactor = 0.005;
numBlocks = 4;
dilationFactor = 1;

net = dlnetwork;
% Define the layers
layers = [
    sequenceInputLayer(3, Normalization="rescale-symmetric", Name="inputaccelerometers")
    convolution1dLayer(filterSize, numFilters, DilationFactor=dilationFactor, Padding="causal", Name="conv1_")
    reluLayer(Name="relu1_")
    fullyConnectedLayer(1, Name="fc")
];

net = addLayers(net,layers)

% Define training options
options = trainingOptions("adam", ...
    MaxEpochs=60, ...
    MiniBatchSize=1, ...
    Plots="training-progress", ...
    Metrics="rmse", ...
    Verbose=0);

% Train the network
net = trainnet(Xtrain', Ytrain', net,"mse",options );

Hope this helps!