Problems with interpretation of training results

[MakkaPakka-sc]

Thank you very much for your research content. I changed the teacher model to GFNet, and the student model used MobileNet V2. I got this result. What is the reason for the mutation when the step is 300? In addition, which indicator represents the test acc of the student model and which represents the acc of the teacher model. Is the calculation of FLOPs integrated in the framework?

[MakkaPakka-sc]

@Zzzzz1 Can you answer my question? Thank you very much

[Zzzzz1]

1. The reason is that the learning rate changes. 2.test_acc represents the acc of the student, the teacher's acc is not plotted. 3. not integrated.

[MakkaPakka-sc]

Thank you for your answer. What may be the reason for the unchanged acc on the later test set?