maybe have a bug

[fangfangzk]

我发现代码中 similarity_map_list里面的4个map，值都相等。原因可能是把中间的feature加入list时没有clone()？

[zqhang]

Hi! Have you checked it for all images or just a single one?

[fangfangzk]

我调试了这里，发现把x[0]添加进out_tokens里面后，然后让 x 继续执行网络，发现out_tokens里面的值也会跟着变

[zqhang]

Thank you, I will check it.

[CHELSEA234]

@fangfangzk @zqhang
I have also noticed this point. May I ask if the performance changes after changing to using x[0]?

[lyy70]

Yes, I also found this BUG, I regenerated the exception graph of the four layers and found that they were consistent. The main reason lies in out_token.append(x[0]). When I changed it to out_tokens.append(copy.deepcopy(x[0])), the patchtoken of the four layers was different, but the exception detection performance decreased.