150 - Warning about keras' data augmentation when working with categorical labels
Image transformation operations, including the ones provided by ImageDataGenerator class from keras, modify numpy arrays based on the user description. The modification of these arrays may yield unintended results as the data may be modified from interpolation. If the user is not careful in spotting these issues the results may turn out to be meaningless. This is especially true when working with categorical labels where each label value needs to be conserved throughout the training and prediction process.
This video walks you through this issue and also offers suggestions on how to handle data augmentation for categorical labels. Custom code for augmentation can be downloaded here: https://github.com/bnsreenu/python_fo...
