How to change color model of all pictures in image datastore?

Hello everybody, I need help with a changing color model of pictures in a Image Datastore.

I've tried to change the color model with a transform database.

clc; clear; close all;
net = resnet101();

%% IMDS
rootFolder = 'cifar100Train';
categories = {'Dolphin','Rocket','Bed','Can','Pear'};
imds.train = imageDatastore(fullfile(rootFolder, categories), 'LabelSource', 'foldernames');
imds.train = splitEachLabel(imds.train, 2 , 'randomize');
auimds.train = augmentedImageDatastore(net.Layers(1).InputSize(1:2), imds.train);

Here comes the problem with transformFcn (function is defined below).

dsnew.train = transform(auimds.train, @transformFcn);

%% Changing layer of a CNN
lgraph = layerGraph(net);
newFCLayer = fullyConnectedLayer(5, 'Name', 'fc5');
lgraph = replaceLayer(lgraph,'fc1000',newFCLayer);
newClassLayer = classificationLayer('Name','new_classoutput');
lgraph = replaceLayer(lgraph,'ClassificationLayer_predictions',newClassLayer);

%% Training options
opts = trainingOptions('sgdm', ...     
    'InitialLearnRate',.001,...
    'MaxEpochs',10, ...
    'MiniBatchSize',64, ...                 
    'Shuffle','every-epoch', ...
    'Plots','training-progress');

CNN = trainNetwork(dsnew.train, lgraph, opts);

%% Changing color model via the transformFcn
function dataOut = transformFcn(dataIn)
dataOut = rgb2hsv(dataIn);
end

For color model change I've used rgb2hsv function. Matlab displays this error message:

Error using trainNetwork (line 170)
Invalid transform function defined on datastore.

Error in transferlearning (line 37)
CNN = trainNetwork(dsnew.train, lgraph, opts);

Caused by:
    Error using nnet.internal.cnn.util.NetworkDataValidator/assertDatastoreHasResponses (line 140)
    Invalid transform function defined on datastore.
        Error using rgb2hsv>parseInputs (line 95)
        MAP must be a Mx3 array.
        

I would love to find any effective solution for changing color model.

Thank you for any answer!

Martin

Answer 1

[Martin Simunsky]

I hope I've figured this problem with Transform datastore:

First error

Error using trainNetwork (line 170)
Invalid transform function defined on datastore.
Error in transferlearning (line 37)
CNN = trainNetwork(dsnew.train, lgraph, opts);
Caused by:
    Error using nnet.internal.cnn.util.NetworkDataValidator/assertDatastoreHasResponses (line 140)
    Invalid transform function defined on datastore.
        Error using rgb2hsv>parseInputs (line 95)
        MAP must be a Mx3 array.

was figured out by using im2uint8 function.

im2uint8(rgb2hsv(dataIn));

Second error

Error using trainNetwork (line 170)
Invalid training data. Responses must be nonempty.

was figured out thanks to this script in transformFcn file:

function dataOut = transformFcn(dataIn)
B = table2cell(dataIn);
NoImg = size(B,1);

for i=1:NoImg
    B{i,1} = im2uint8(rgb2hsv(B{i,1}));
end
dataOut = cell2table(B);
end

In main file:

imds.train = imageDatastore(fullfile(trainDirectoryName, categories), 'LabelSource', 'foldernames');
imds.train = splitEachLabel(imds.train,5, 'randomize');
auimds.train = augmentedImageDatastore(sizeImg, imds.train);
dsnew.train = transform(auimds.train, @transformFcn);
dsnew.train.UnderlyingDatastore.MiniBatchSize = 10;

ReadFcn takes only one image, instead Transform datastore takes whole batch of images (size of this batch is defined in the last line: dsnew.train.UnderlyingDatastore.MiniBatchSize; default is 128).

Hope this helps :-)

Answer 2

[Roshni Garnayak]

Instead of using ‘transformFcn’, you can use the ‘ReadFcn’ parameter in the ImageDatastore object. You can define the colour space transformation in the @customreader.

For more information on how to do it, please refer the Properties section in the following link:

https://www.mathworks.com/help/matlab/ref/matlab.io.datastore.imagedatastore.html

[Martin Simunsky]

ReadFcn parameter works

I've tried this option as well. But they say this:

"Using ReadFcn to transform or pre-process 2-D images is not recommended. For file formats recognized by imformats, specifying ReadFcn slows down the performance of imageDatastore. For more efficient ways to transform and pre-process images, see Preprocess Images for Deep Learning (Deep Learning Toolbox)." (1)

Howewer, after few attempts it works:

...
imds.train = imageDatastore(fullfile(trainDirectoryName, categories), 'LabelSource', 'foldernames');
imds.train = splitEachLabel(imds.train,5, 'randomize');
imds.train.ReadFcn = @transformFcn;
...

transformFcn file:

function dataOut = transformFcn(dataIn)
dataOut = imread(dataIn); 
dataOut = im2uint8(rgb2hsv(dataOut));
end

transformFcn displays error

If I used transformFcn instead of ReadFcn in this way:

...
imds.train = imageDatastore(fullfile(trainDirectoryName, categories), 'LabelSource', 'foldernames');
imds.train = splitEachLabel(imds.train,5, 'randomize');
% imds.train.ReadFcn = @transformFcn;
dsnew.train = transform(imds.train, @transformFcn);
...
    

transformFcn file:

function dataOut = transformFcn(dataIn)
% dataOut = imread(dataIn); 
dataOut = im2uint8(rgb2hsv(dataIn));
end

Matlab displays this error message in the beginning of training:

Error using trainNetwork (line 170)
Invalid training data. Responses must be nonempty.

Can anyone help me, please?