trainNetwork - Number of observations in X and Y disagree

Hi,

I'm training a network that has an imageInputLayer with image size of [498,13]. I have 1600 instances of these 498x13 images in my training data set.

The dimensions of the variables I'm passing to the trainNetwork function are the following:

Xtrain: 498x13x1600

Labels: 1600x1

These dimensions are as reccomended in the documentation:

Dimensions for X: hxwxN & Dimensions for Y: Nx1 are as specified in the answer to a question on this error asked previously.

But I'm still getting the error "Number of observations in X and Y disagree" when I call trainNetwork(Xtrain, Labels, layers, options)

Do not understand why I'm getting the error.

Thanks,

Nikhil

[Image Analyst]

Just to be clear, replace this line

trainNetwork(Xtrain, Labels, layers, options)

with these lines

whos Xtrain
whos Labels
trainNetwork(Xtrain, Labels, layers, options)

and tell us what you see in the command window.

[Nikhil]

Here's what I get I run the code below:

whos XtrainShaped
whos TTrain
trainedNet = trainNetwork(XtrainShaped,TTrain,layers,options);

Answer 1

[Rahul]

"XtrainShaped" dimension should be [img_height, img_width, img_planes, num_samples]. "img_planes" denotes whether is image is RGB (img_planes = 3) or grayscale (img_planes = 1). In your dataset, the network is thinking the number of planes are 1600 with only one sample whereas length of the Labels categorical vector is 1600. Hence, you are receiving the error. You can reshape your "XtrainShaped" matrix using the link hyperlink given.

The shape of your XtrainShaped should be . The 1 represents the number of planes in an image denoting whether the image is grayscale or RGB. Please refer the below code for more information. This way, you will not receive any syntax error.

Important Note: The accuracy and other metrics depend on your CNN model, training options, and other hyperparameters. Below code is just for demo purpose to avoid the error that you are receiving.

imdsTrain = rand(498, 13, 1600); % generating random samples (your dataset)

% RESHAPE YOUR DATA
imdsTrain_reshaped = reshape(imdsTrain, [498, 13, 1, 1600]); % reshaping to [height, width , num_planes, num_samples]


labels = randi([0,9],1600, 1); % generating vector having 10 classes
labels = categorical(labels); % converting to categorical vector

[img_height, img_width , num_planes, num_samples] = size(imdsTrain_reshaped);
num_classes = 10; 

% CNN layers
layers = [ ...
    imageInputLayer([img_height, img_width , num_planes]) % input layer
    convolution2dLayer(3,20) % 2D convolutional layer
    reluLayer % activation function
    maxPooling2dLayer(2,'Stride',2) % 2D max-pooling layer
    fullyConnectedLayer(num_classes) % Fully connected later with output neurons 10
    softmaxLayer % activation function
    classificationLayer]; % Classification layer

% training options
options = trainingOptions('sgdm', ...
    'MaxEpochs',5,...
    'InitialLearnRate',1e-4, ...
    'Verbose',true, ...
    'Plots','training-progress');

% training the network
net = trainNetwork(imdsTrain_reshaped, labels, layers, options);

[Nikhil]

Thank you for explaining how the image dimension should be specified for imageInputLayer.

I reshaped the input to be a 498x13x1x1600 vector, but still getting the same error message.

whos XtrainShaped
whos TTrain
trainedNet = trainNetwork(XtrainShaped,TTrain,layers,options);

[Rahul]

Can we have a screen sharing session as per your convenience? Suggest a suitable time. We can google meet.

[Rahul]

Check "imageinputlayer" in "layers" and check the dimesions.

imageInputLayer([img_height, img_width , num_planes]) % input layer

[Nikhil]

I'm setting as below:

[img_height, img_width , num_planes, num_samples] = size(XtrainShaped)
%Define layers of the network
layers = [
    imageInputLayer([img_height, img_width , num_planes]) % input layer
   
I get the following 

I get the following values for the image dimensions:

But, still getting the error when I run the code below:

whos XtrainShaped
whos TTrain
trainedNet = trainNetwork(XtrainShaped,TTrain,layers,options);

[Rahul]

Please contact MathWorks technical support to resolve this issue.

[Rahul]

I am using MATLAB R2022a and I did not encounter any such error. If possible, please upgrade the MATLAB software to R2022a or later to avoid this error. If this is not possible, please let me know the MATLAB release that you are using and I will try to reproduce the issue at my end. We can also schedule a screen sharing session at your availability.

[Nikhil]

I'm also using R2022a.

Please let me know your availability for a screen sharing session this weekend.

Thanks, Nikhil

[Rahul]

Saturday 22nd, 2022 after 05:30 PM IST or Sunday 23rd, 2022 anytime. Please suggest a suitable time as per your convenience and I shall mail you the google meet link details.

[Nikhil]

Does 8:00 PM or 8:30 PM IST on Sunday 23rd, work?

Thanks, Nikhil

[Rahul]

Sure 8:00 PM IST works. I have mailed you the google meet link details on your gmail account. Feel free to reply.

[PABLO MARTIN PUERTO]

Did you get the solution for that? I am currently having the same problem. Thanks

Answer 2

[伟 王]

Did you solve this problem, I am also experiencing the same situation