Some questions about released code

[YeRen123455]

@jbeomlee93 ! Sorry for disturbing you again. I still have two questions about the released code. (1) In obtain_CAM_masking_super_pixel.py. Since you have used grad-cam to generate the class activation map(i.e., CAM), why don't you use resnet50.py with grad-cam to generate outputs. Actually, you used resnet50_cam.py with grad-cam to generate the outputs.

(2) Can you share the code of "SEAM+AdvCAM" with me. I try to reproduce it by myself but the performance is not good as yours. My email address is liboyang20@nudt.edu.cn

[YeRen123455]

Hi @jbeomlee93 ! I tried to reproduce "SEAM+AdvCAM" by myself but only achieved 50.96mIoU. I think the difference (50.96 vs 58.6) was caused by my wrong experimental setting. I used the following setting for "SEAM+AdvCAM": (1) Since the output of SEAM has 21 classes, I removed the class of "background" to attack the images. (2) I used the layer "f9" in SEAM as the target layer to generate CAM. Maybe you have used some special settings for it. I want to reproduce it.

[jbeomlee93]

Hi @YeRen123455 , sorry for the late reply.

(1) "resnet50.py" just contains the definitions of layers of resnet50, and actual architectures for classification and CAM are included in "resnet50_cam.py". I followed the default configuration of IRN (https://github.com/jiwoon-ahn/irn).

(2) On the "SEAM+AdvCAM", you can consider these issues: https://github.com/jbeomlee93/AdvCAM/issues/7#issuecomment-825309703, https://github.com/jbeomlee93/AdvCAM/issues/8#issuecomment-842050602. I think you should change the values of hyper-parameters.

Thanks.

[YeRen123455]

@jbeomlee93 Thanks for you reply. I will try it again.

[YeRen123455]

@jbeomlee93 Since the CAM output of SEAM has 21 classes (20 classes + background), the default class label for your adv_cam is 20. So you also attck the background On the "SEAM+AdvCAM"?

[jbeomlee93]

No, I just apply adversarial climbing only for the (ground-truth) foreground labels.

[YeRen123455]

Hi @jbeomlee93 Thanks for your reply. I have followed your suggestion in #7 (comment), #8 (comment) and retrained the "SEAM+Adv_cam". But the performance is even worse. I have changed 4 places of your "obtain_CAM_masking.py" to reproduce the "SEAM+Adv_cam". The changes are as follows:

[ 1] I followed your reply in combine with seam Can you provide your trained weights for MDvsFA-cGAN? #7 (comment) and set the masking threshold $T$, regularization item $\lambda$ as 8 and 2, respectively. That is:

       parser.add_argument("--AD_coeff",    default=2, type=int)
       parser.add_argument("--score_th",    default=8, type=float)     in  "obtain_CAM_masking.py"

[ 2] I tried to obey your suggestion in Question about Table 4. 文章中的错误 #8 (comment), but if the adversarial climbing was done on logit, GAP(cam), before up-sampled and before PCM module. It means I should output the cam value in forward function defined in "resnet38_SEAM.py". That is：

      def forward(self, x, separate=False):
      N, C, H, W = x.size()
      d = super().forward_as_dict(x)
      cam = self.fc8(self.dropout7(d['conv6']))
      if separate:
      return cam in "resnet38_SEAM.py"

[ 3] At the same time, I should also accumulate our localization maps (in Equation 4) using CAM after PCM module. It means I should change the code of "resnet38_SEAM.py" as:

  def forward(self, x, separate=False):
  N, C, H, W = x.size()
  d = super().forward_as_dict(x)
  cam = self.fc8(self.dropout7(d['conv6']))
  ............
  ............
  cam_rv_1 = self.PCM(cam_d_norm, f)
  if separate:
  return cam_rv_1 in "resnet38_SEAM.py"

However, the step 2 and 3 are contradictory. I can only output cam or cam_rv_1. Otherwise I can not get the value of "regions" in your "obtain_CAM_masking.py" by grad-cam (that is because I can not get the model's gradient if I output both cam and cam_rv_1).

[ 4] Since the CAM output of SEAM has 21 classes (20 classes + background), the default class label for your adv_cam is 20. I change the "gradCAM.py" as follows:

  self.logits = self. model(image, separate= True)[ :, 1:, :, :] in "gradCAM.py"

The above changes can not make me to successfully reproduce SA. Could you please help me to check these changes. I really want to follow your work! Thanks a lot!