augmentation of pretrained network reducing accuracy

Hi, image augmentation using pretrained network is reducing accuracy of network. The file 1060.png is with dataaugmentation which reducces the accuracy as compared to 1061.png which has better accuracy without dataaugmentation.

what can be the reason, I read that augmentation increases accuracy.

Answer 1

[Chunru]

Generally speaking, training with data augmentation will lead to poorer traing accuracy, since the model needs to fit more training data. Data augmentation acts as a regularize and helps reduce overfitting. Therefore, it generally produce better generalization result. If you have different test/validation dataset, it usually improve the test/validation accuracy instead of training accuracy (as you have shown in the attached figure). Try to include some test/validation data in your training to see if the test/validation accuracy improves with data augumentation.