Dimensions of arrays not consistent error with image segmentation network

I am getting an error on line 80 of my code (when I try to train the network) where it says that the dimensions of the arrays being concatenated are not consistent. After looking into the problem a bit I double checked and all of my input images are 1024x1024 so I'm not sure where the problem is coming from.

clear all;
clc;
 
dataSetDir = 'C:\Users\jacob\Documents\School\Image Segmentation Training Data';
imageDir = fullfile(dataSetDir, 'Cores');
labelDir = fullfile(dataSetDir,'Labels Despeckeled');

imds = imageDatastore(imageDir);

classNames = ["background","core"];
labelIDs   = [255 0];
pxds = pixelLabelDatastore(labelDir,classNames,labelIDs);
 
% I = read(imds);
% C = read(pxds);
% 
% I = imresize(I,5);
% L = imresize(uint8(C{1}),5);
% imshowpair(I,L,'montage')
 
numFilters = 64;
filterSize = 3;
numClasses = 2;
 
inputSize = [1024,1024,3];
imgLayer = imageInputLayer(inputSize); 
 
conv = convolution2dLayer(filterSize,numFilters,'Padding',1);
relu = reluLayer();

poolSize = 2;
maxPoolDownsample2x = maxPooling2dLayer(poolSize,'Stride',2);

downsamplingLayers = [
   conv
   relu
   maxPoolDownsample2x
   conv
   relu
   maxPoolDownsample2x
   conv
   relu
   maxPoolDownsample2x
   ];

filterSize = 4;
transposedConvUpsample2x = transposedConv2dLayer(4,numFilters,'Stride',2,'Cropping',1);

upsamplingLayers = [
   transposedConvUpsample2x
   relu
   transposedConvUpsample2x
   relu
   transposedConvUpsample2x
   relu
   ];

conv1x1 = convolution2dLayer(1,numClasses);

finalLayers = [
   conv1x1
   softmaxLayer()
   pixelClassificationLayer()
   ];

layers = [
   imgLayer    
   downsamplingLayers
   upsamplingLayers
   finalLayers
   ];

opts = trainingOptions('sgdm', ...
    'InitialLearnRate',1e-3, ...
    'MaxEpochs',15, ...
    'MiniBatchSize',64, ...
    Plots="training-progress");

trainingData = combine(imds,pxds);
net = trainNetwork(trainingData,layers,opts);

testImage = imread('C:\Users\jacob\Documents\School\Image Segmentation Training Data\Cores\00.1nM Dox 231-001.png');
imshow(testImage)

C = semanticseg(testImage,net);
B = labeloverlay(testImage,C);
imshow(B)

[mohd akmal masud]

input size are not match with groundtruth size.

[Jacob Heiss]

All of the pixel label images are also 1024x1024.

[Pratham Shah]

Hi!

Please share the screenshot of the error. Without your dataset we can't regenerate the error.

[Jacob Heiss]

Here is a screenshot of the error. Thanks.

[Jacob Heiss]

Here is also an example of an image and its label

Answer 1

[Jacob Heiss]

I figured out the issue, the pixel label images had an invert LUT on them which was causing the problem.

Answer 2

[Vinayak Agrawal]

Hi Jacob,

Based on the code snippet you provided, to fix this error, you should do following change

'Plots','training-progress');

By changing this, MATLAB should be able to correctly generate the desired training progress plots during the training process.

You can see this way of generating plots from this documentation(Train deep learning neural network - MATLAB trainNetwork - MathWorks India)

Here is the screenshot from the same document from where the changes are taken

Hope it helps!

Please accept the answer if it is working so that others can get benefitted.

[Jacob Heiss]

That change did not fix the issue. The previous syntax still started to generate the plot.

[Vinayak Agrawal]

try changing it to lowercase

'plots','training progress');

[Jacob Heiss]

I don't think that the issue is with the syntax of that part. The code still has the error even when that part is completely removed.