If I understand your use case correctly, you are not allowed to use data augmentation during the training, but would like to achieve a similar performance with a static dataset?
If so, I wouldn’t know of a good technique to act as a replacement for data augmentation. Maybe try to increase the regularization through dropout layers, weight decay etc.
1 Like