Hello Professor,

As it's an augmented dataset, I am assuming you have already balanced the dataset. Since the learning rate is also not so high, i would suggest you experiment with an increase in the number of layers in model, (eg, using a ResNet architecture which has more flip-flops than a VGG) , else changing the hyperparameters and monitor the epochs and terminate the model as soon as you notice overfitting.

However, it still depends on the dataset features respectively.

Thank you.