Why training/validation accuracy and loss are a little bit worst with data augmentation

Question

I am using transfer learning technique on Alexnet and other models along with image augmentation using pytorch lightning. The f1 scores look good, as I would expect, the val