sungraepark / adversarial-dropout Goto Github PK

Hi, I have tested the dataset of cifar10 based on your code (Adversarial-Dropout), but I cannot achieve the results showed in your paper. My experimental results are as follows:

Method | labels | paper | My experiment
VAD(KL)_VAT | 50000 | 4.40 | 9.86
VAD(KL) | 50000 | 5.27 | 17.12
∏ model | 4000 | 12.62 | 28.17
VadD(KL) | 4000 | 11.68 | 15.43
VadD(QE) | 4000 | 11.32 | 18.07
VAD(KL)_VAT | 4000 | 10.07 | 12.73

From the above table, my results are worse than yours. I used your codes and your parameters, I don't know why? can you help me? Thanks.

Hi, does adversarial dropout have input scaling?

Sorry in case I miss this detail is in code or in the paper.