Regarding the code

[laukun]

Thank author for the code. I have two small questions:

1. regarding this small block of code below to obtain cam: def forward_cam(self, x): x = super().forward(x) x = self.fc8(x) x = F.relu(x) x = torch.sqrt(x) return x what's x=torch.sqrt(x) for?
2. regarding the pretrained caffe weghts, i.e. vgg16_20M.prototxt, on what dataset was it trained on? were you meant to use the deeplab_LargeFOV model of vgg16 version as the network to compute CAM, but why setting fc6_dilation=1 not 12 as in deeplab v1 paper?

[jiwoon-ahn]

• VGG16_20M model was trained on ImageNet as deeplab v1 paper says.
• I just wanted to adopt DeepLab in CAM and AffinityNet because it will be used for semantic segmentation anyway.
• I think other questions are highly related to the above. To convert DeepLab as a CAM network, I followed https://arxiv.org/pdf/1701.08261.pdf. (which includes removing last max polling layer, adjusting dilation rate, etc.)
• Then, I found that as the last maxpooling layer was removed, the varaince of the activation map has become different which can be simply compensated by x = torch.sqrt(x). (This is not neccessary when tyring with other backbone networks such as VGG-GAP https://github.com/metalbubble/CAM)