SMOTE/GAN for image data classification in FastAI

Hi Everybody,

I am working on an image classification problem where I have unbalanced classes. I’ve 4 classes as following:
Class A and B ~ 600 Examples each
Class C ~ 500 Examples
Class D ~ 100 Examples
I’m using OverSamplingCallback but I would also like to try SMOTE/GAN and see how they perform.
I’ve a couple of questions:

  1. How can I oversample my images using SMOTE in FastAI?

  2. How can I oversample my images using GAN in FastAI as mentioned in this paper?

Any help would be highly appreciated.

Apart from these two questions please feel free to add your 2 cents about dealing with imbalanced classes, if you would like to.


1 Like