Customized data augmentation

[hamidfarhid]

Hi,
I was looking for customized data augmentation for image data set that I have and found this:

How to augment the minority class only in an unbalanced dataset

I have an unbalanced image dataset with the positive class being 1/10 of the entire dataset. Classification models trained on this dataset tend to be biased toward the majority class (small false negative rate and bigger false positive rate). I would like to do some augmentation only on the minority class to deal with this. The way I understand, using transforms (random rotation, etc.) when creating torchvision.dataset will augment all of the data so the imbalance remains the same. What is the …

This would be great to start as @ptrblck pointed. However, I am looking to the clean method to augment data to the certain number of images, let’s say 5000 images per each class. I have various distribution of data across all classes, from 200 to 1000, so I have to augment some classes more than others. Is there any way to do so?

Thanks

[Stardust (Naman Rastogi)]

First of all, we can select the classes that we have to focus on (minority classes for imbalanced dataset), then we can define a list of transformations for the selected classes and give the probability values for each individual transformations, like for classes that are very rare, we should augment them for 90 percent (for example) of the cases, and others that are in majority, augment them for only 10 percent (for example) of the cases, and so on. In this way, in the custom dataset class that we created, we can iterate on the classes and pass the predefined transformations with certain probability values, and apply on those classes/labels.