C-RNN regression error (Array inputs have incompatible channel dimensions)

Hi.

I am writing a C-RNN regression learning code. The loaded "paddedData2.mat" file is saved as paddedData, and it is stored as an N X 3 cell, as shown in the attached image. The input matrix used for training is the 3rd column of paddedData, which is [440 5] double, and the regression variable is the values in the 1st column. With this, I plan to create features of size [436 1] using two [3 3] kernels of convolution and train them using LSTM. The code is as follows. When running this code, the training window pops up, and at the same time, an error message "Array inputs have incompatible channel dimensions." occurs. Is there a solution?

clc;
clear all;
load("paddedData2.mat","-mat")
XTrain = paddedData(:,3);
YTrain1 = cell2mat(paddedData(:,1));

layers = [
    sequenceInputLayer([440 5 1])
    convolution2dLayer(3,8)
    batchNormalizationLayer
    reluLayer
    convolution2dLayer(3,8)
    batchNormalizationLayer
    reluLayer
    flattenLayer
    fullyConnectedLayer(100)
    lstmLayer(100,'OutputMode','last')
    fullyConnectedLayer(1)
    regressionLayer];

options = trainingOptions('adam', ...
    'MaxEpochs',2000, ...
    'MiniBatchSize',100, ...
    'Plots','training-progress');

net = trainNetwork(XTrain, YTrain1, layers, options);

Answer 1

[Angelo Yeo]

It might be a bug in R2022a, but I'm not sure where exactly the error comes from because it uses a "builtin" function. However, it looks like the issue is fixed in a recent release. You can see that there is no issue in R2024a.

Please update to the most recent version of MATLAB and run the code.

clear;
paddedData = cell(10, 3); % Intentionally changed the #observations from 900 to 10 for reproduction purpose
paddedData(:,1) = cellfun(@(x) 405, paddedData(:,1), 'UniformOutput', false);
paddedData(:,2) = cellfun(@(x) 10, paddedData(:,2), 'UniformOutput', false);
paddedData(:,3) = cellfun(@(x) rand(440, 5), paddedData(:,3), 'UniformOutput', false)

paddedData = 10x3 cell array
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}
    {[405]}    {[10]}    {440x5 double}

XTrain = paddedData(:,3);
YTrain1 = cell2mat(paddedData(:,1));

layers = [
    sequenceInputLayer([440 5 1])
    convolution2dLayer(3,8)
    batchNormalizationLayer
    reluLayer
    convolution2dLayer(3,8)
    batchNormalizationLayer
    reluLayer
    flattenLayer
    fullyConnectedLayer(100)
    lstmLayer(100,'OutputMode','last')
    fullyConnectedLayer(1)
    regressionLayer];

options = trainingOptions('adam', ...
    'MaxEpochs',10, ... % Intentionally changed the max epochs for reproduction purpose
    'MiniBatchSize',100, ...
    'Plots','none'); % Intentionally changed 'none' for reproduction purpose

net = trainNetwork(XTrain, YTrain1, layers, options);

Training on single CPU.
|========================================================================================|
|  Epoch  |  Iteration  |  Time Elapsed  |  Mini-batch  |  Mini-batch  |  Base Learning  |
|         |             |   (hh:mm:ss)   |     RMSE     |     Loss     |      Rate       |
|========================================================================================|
|       1 |           1 |       00:00:00 |       405.00 |      82012.9 |          0.0010 |
|      10 |          10 |       00:00:00 |       397.93 |      79175.6 |          0.0010 |
|========================================================================================|
Training finished: Max epochs completed.