Error forming mini-batch for network input "input". Data interpreted with format "SSCB". To specify a different format, use the InputDataFormats option.

I am working on a binary classification of images.the code i am using is as follows.

In short, the 'data' file contains an image as 128*128*2 in its cell. 128 is width, 128 is height and 2 is channel (gmti is a name for channel 1, ref is a name for channel 2).

I have 290 images in total so 'data' is a 290*1 cell. 290 as batch, 128*128 as spece, 2 as channel. (SSCB)

For labels, I simply labeled then as double 0 or 1, and used num2str and converted it into categorical.

the error message is as such.

trainnet:

Error forming mini-batch for network input "input". Data interpreted with format "SSCB". To specify a different format, use the InputDataFormats option.

net_trained = trainnet(data,labels_cat, net,'crossentropy',options);

Cause:

Cell array input supports sequence data only.

I tried changing input layer as sequence and making data format as 128*128*2*290 double. Did not work.

Exactly what is the proper data format should i use?

(Due to file size constraint, data_2 file consists of ones and zeros)

----------------------------------------------------------------------------------------------------------

clear 
close all

load data_2.mat 
load labels_cat.mat
% imds = imageDatastore(data, 'Labels',labels_cat);
%%
imageSize = [128 128 2];
numClasses = 2;

stackDepth = [3 4 23 3];
numFilters = [64 128 256 512];

net = resnetNetwork(imageSize,numClasses, ...
    StackDepth=stackDepth, ...
    NumFilters=numFilters);
% analyzeNetwork(net)
% net = replaceLayer(net, 'input', sequenceInputLayer(imageSize));

options = trainingOptions("sgdm", ...
    InitialLearnRate=0.01, ...
    MaxEpochs=100, ...
    Shuffle="every-epoch", ...
    Plots="training-progress", ...
    InputDataFormats = "SSCB", ...
    Verbose=false);
    % InputDataFormats = "SSCB", ...
%%
%%
net_trained = trainnet(data,labels_cat, net,'crossentropy',options);

[Matt J]

The code that sets up your data doesn't include enough to be run:

for i = 1 : size(gmti, 1)
    data{i,1}(:,:,1) = [gmti{i}];
    data{i,1}(:,:,2) = [ref{i}];
end

Unrecognized function or variable 'gmti'.

nor is it necessary to give us this part. Just attach a .mat file with selections of data and labels_cat.

[chanul]

mat files are upload with runnable code

Answer 1

[Gayathri]

Hi @chanul,

I was able to reproduce the error mentioned. Before processing image data with networks like ResNet, it's essential to first convert cell-formatted data into an array.

data1 = zeros(128, 128, 2, 290);
for i = 1:290
    data1(:, :, :, i) = data{i};
end

And then when passing the input data to the “trainnet” function for training, the following change can be accommodated.

net_trained = trainnet({data1,labels_cat}, net,'crossentropy',options);

I hope the approach mentioned above will help resolve the issue.