How to draw attention weight in ROIGather

[GreatZeZuo]

Thanks for your work, can you please tell me how to draw figure 5 in the paper. How to draw attention weight on the original figure ? I would appreciate it if you could respond in a timely manner.

[Napier7]

I tried to implement an attention visualization function myself, but it didn't seem to be effective. Did you solve it?

[Napier7]

I tried to implement an attention visualization function myself, but it didn't seem to be effective. Did you solve it?

Then I tried adjusting the code again, and now it seems to work better. You can add code starting from line 221, /libs/models/layers/attentions.py：（please ignore my code quality）

# add 'layer_index' parameter passing on line 148
if layer_index == 2: 
  import numpy as np
  from torchvision import transforms
  import cv2
  B, Np = sim_map.shape[:2]
  # cancel the code 'x=F.interpolate(x,self.size)' on line 262
  att_map = sim_map.reshape(B, Np, 40, 100).mean(dim=1).unsqueeze(1)
  max_val = att_map.reshape(B, 1, -1).max(dim=-1).values 
  min_val = att_map.reshape(B, 1, -1).min(dim=-1).values 
  att_map = torch.div(att_map - min_val, max_val - min_val)
  trans = transforms.Resize([540, 960])
  att_map = trans(att_map).squeeze(0).permute(1,2,0) * 255
  att_map = np.uint8(np.array(att_map.cpu()))
  att_map = cv2.applyColorMap(att_map, cv2.COLORMAP_JET)
  cv2.imwrite('attention_map.png', att_map)
  import pdb; pdb.set_trace()

[zoro-dong]

I tried to implement an attention visualization function myself, but it didn't seem to be effective. Did you solve it?

Then I tried adjusting the code again, and now it seems to work better. You can add code starting from line 221, /libs/models/layers/attentions.py：（please ignore my code quality）

# add 'layer_index' parameter passing on line 148
if layer_index == 2: 
  import numpy as np
  from torchvision import transforms
  import cv2
  B, Np = sim_map.shape[:2]
  # cancel the code 'x=F.interpolate(x,self.size)' on line 262
  att_map = sim_map.reshape(B, Np, 40, 100).mean(dim=1).unsqueeze(1)
  max_val = att_map.reshape(B, 1, -1).max(dim=-1).values 
  min_val = att_map.reshape(B, 1, -1).min(dim=-1).values 
  att_map = torch.div(att_map - min_val, max_val - min_val)
  trans = transforms.Resize([540, 960])
  att_map = trans(att_map).squeeze(0).permute(1,2,0) * 255
  att_map = np.uint8(np.array(att_map.cpu()))
  att_map = cv2.applyColorMap(att_map, cv2.COLORMAP_JET)
  cv2.imwrite('attention_map.png', att_map)
  import pdb; pdb.set_trace()

Can you tell me more details? I would be very grateful.Extremely urgent. My contact information is [email protected].

[Napier7]

I tried to implement an attention visualization function myself, but it didn't seem to be effective. Did you solve it?

Then I tried adjusting the code again, and now it seems to work better. You can add code starting from line 221, /libs/models/layers/attentions.py：（please ignore my code quality）

# add 'layer_index' parameter passing on line 148
if layer_index == 2: 
  import numpy as np
  from torchvision import transforms
  import cv2
  B, Np = sim_map.shape[:2]
  # cancel the code 'x=F.interpolate(x,self.size)' on line 262
  att_map = sim_map.reshape(B, Np, 40, 100).mean(dim=1).unsqueeze(1)
  max_val = att_map.reshape(B, 1, -1).max(dim=-1).values 
  min_val = att_map.reshape(B, 1, -1).min(dim=-1).values 
  att_map = torch.div(att_map - min_val, max_val - min_val)
  trans = transforms.Resize([540, 960])
  att_map = trans(att_map).squeeze(0).permute(1,2,0) * 255
  att_map = np.uint8(np.array(att_map.cpu()))
  att_map = cv2.applyColorMap(att_map, cv2.COLORMAP_JET)
  cv2.imwrite('attention_map.png', att_map)
  import pdb; pdb.set_trace()

Can you tell me more details? I would be very grateful.Extremely urgent. My contact information is [email protected].

That is all the details.