Hello,liruihui.Thank you for you contribution!
I find in your article, you split into two parts.One part is update the classifier, the other one is update augmentor.Is the reason why you implemented end-to-end because you used the loss function of Equation6 to update the classifier?