Last dimension disappears when I use an arrayDatastore

Hello everyone,

I was using a 4-d array of shape 8x234x1x1000 to train a ResNet, 8x234 being the size of my matrix, 1 being the number of channels and 1000 the number of points in my training set.

It worked well, however, I now want to use a multiple input neural network and hence need to use an arrayDatastore (I think?)

I specify : XTrain_features = arrayDatastore(features_mat(:,:,:,trainInd), "IterationDimension",4);

However, read(XTrain_features) gives :

1×1 cell array

{8×234 double}

And my neural network which input layer is imageInputLayer([8 234 1]) now gives an error saying I have no channel dimension.

Could someone please give me a workaround?

Thanks!

[Stephen23]

https://www.mathworks.com/matlabcentral/answers/21992-singleton-dimention-as-last-dimension-in-matrix#answer_28947

https://www.mathworks.com/matlabcentral/answers/1971659-how-to-expand-dimension-of-a-2d-array

https://www.mathworks.com/matlabcentral/answers/229956-how-can-i-reshape-a-20370x1-matrix-into-2037x10x1

https://www.mathworks.com/matlabcentral/answers/284047-applying-mean-to-multidimensional-matrix-leads-to-disappearance-of-dimension

It would be a very badly-written tool that did not accept a dimension of size one.

Answer 1

[Aneela]

Hi @Julien,

MATLAB implicitly performs squeezing action on the arrays. It removes singleton dimensions from an array. So, if an array has a size of 2x3x1, MATLAB will often display or treat it as a 2x3 array. This is called “Trailing Singleton Dimensions”.

Please refer to the below MATLAB Answers from the community for more information on “Trailing Singleton Dimensions”:

• https://www.mathworks.com/matlabcentral/answers/1971659-how-to-expand-dimension-of-a-2d-array
• https://www.mathworks.com/matlabcentral/answers/229956-how-can-i-reshape-a-20370x1-matrix-into-2037x10x1

Furthermore, for dlnetworks like “resnet18”, “resnet101”, “resnet50”, it is recommended to use “trainnet” for training.

Please refer to the following MathWorks documentation for more details on various datastores supported by “trainnet” for different input data.

https://www.mathworks.com/help/deeplearning/ref/trainnet.html