MATLAB: Data Augmentation not working

data augmentationimage augmentationMATLABneural networktraining

Hi,

I'm trying to do data augmentation using 'ImageDataAugmenter' on a PixelLabelImageDatastore but the rate of my training is not decreasing (as it should if more images are added).

I'm not getting any error but the my training progress shows the same number of iterations and epochs as my non-augmented dataset.

This is my code:

imageAugmenter = imageDataAugmenter( ...
    'RandRotation',[-10,10], ...
    'RandXTranslation',[-10 10], ...
    'RandYTranslation',[-10 10])
augimds = pixelLabelImageDatastore(imds,pxds,'DataAugmentation',imageAugmenter,'OutputSize',imageSize)
trainedNet_aug = trainNetwork(augimds,lgraph,options)

Does Data augmentation add new images to the datastore? Because it still shows my original 300 images in the augimds variable.

Best Answer

As per the MATLAB documentation for “imageDataAugmenter”, the augmented images are not saved to memory. So, the size of “augimds” would effectively remain the same in MATLAB workspace. But these images are used while training.

You may refer to the following link on “imageDataAugmenter” for any clarifications.

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

Related Solutions

MATLAB: Is there a way to visualize augmented image data created by imageDataAugmenter()

For R2018a.

In R2018a, the ‘augmentedImageSource’ has been replaced by ‘augmentedImageDatastore’ and the ‘preview’ method can be used to preview the data as follows:

augimds = augmentedImageDatastore(…)
ims = augimds.preview();
montage(ims{1:6,1}) % to visualize 6 images

'preview' does not reset the datastore.

In alternative, the method ‘read’ can be used to view the next batch of data as shown below:

augimds = augmentedImageDatastore(…);
numBatches = ceil(augimds.NumObservations / augimds.MiniBatchSize);
for i = 1:numBatches
    ims = augimds.read();
    montage(ims{:,1});
    pause;
end
augimds.reset();

____________________________________________________________________________________________________________________

For R2017b:

The built-in functionality to do this is not available on Neural Network Toolbox (R2017b).

1) One of the possible workarounds is to make use of augmentedImageSource's hidden methods and properties for iterating over the data. For example:

[XTrain,YTrain] = digitTrain4DArrayData;
imageAugmenter = imageDataAugmenter('RandRotation',[-180 180],'RandXReflection',true,'RandYReflection',true);
imageSize = [28 28 1];
datasource = augmentedImageSource(imageSize,XTrain,YTrain,'DataAugmentation',imageAugmenter);
datasource.MiniBatchSize = 16;
datasource.reset();
numBatches = ceil(datasource.NumberOfObservations / datasource.MiniBatchSize);
for i = 1:numBatches
    ims = datasource.nextBatch();
    montage( cat(4, ims) );
    pause;
end

This will show a 4x4 montage of augmented data. You can cycle through all the data by pressing any key on the keybord.

2) Another workaround is to set a breakpoint on line 484 of imageDataAugmenter and then visualize variable "B" by using imshow().

3) A third way would be to create a new layer just for image visualization. This layer could be placed after the imageInputLayer and all it would do is display the input images (again using 'montage' for example) during training. You would only need to edit the 'predict' method with the code to display the incoming images (X).

The documentation on how to create a new deep learning layer can be found here:

https://www.mathworks.com/help/nnet/ug/define-custom-deep-learning-layers.html

MATLAB: Data augmentation in CNN

The following are few suggestions:

Make sure that all the classes have equal number of observations.
Check how trainNetwork uses an augmented image datastore to transform training data for each epoch: Augment Images for Training with Random Geometric Transformations. Then try training the network with augmentedImageDatastore for more epochs.
Try changing the network architecture itself if there is no improvement in the accuracy when augmentedImageDatastore is used. You can refer to Choose Network Architecture.
Try Using dropout layers & increasing global L2 regularization factor in new architecture. For more information, see dropoutLayer & 'L2Regularization' option in trainingOptions.

Best Answer

Related Solutions

MATLAB: Is there a way to visualize augmented image data created by imageDataAugmenter()

MATLAB: Data augmentation in CNN

Related Question