I'm training a convolutional neural networks for image segmentation. In training data preprocessing i'm applying some data augmentation to change luminosity of images. I'm using tensorflow random_brightness and random_contrast, fixing the seed i can obtain the same augmentation to the same image at each iteration on the dataset. My question is: data augmentation should remain fixed with the augmented images that don't change during all the training process? Or i can change randomly the images during epochs (without fixing the seed)?