Hey authors!!! Thank you for implementation of IEN in contract to steerable CNNs. My question is regarding the equivarince loss. I understand that one term which goes into the loss, is the GroupPooled output from the layer. But, what exactly is the other term?
Let's say x1 is an image from the dataset X, then the GroupPooled output from the layer i is the first term. How exactly is the other term here defined? lines found here
temp_rot = torch.cat((temp_rot, augmented_temp.unsqueeze(0)),0)
loss2.append(equivariance_loss(temparr[j][:tbs],temp_rot))
Hey authors!!! Thank you for implementation of
IENin contract to steerable CNNs. My question is regarding theequivarince loss. I understand that one term which goes into the loss, is theGroupPooledoutput from the layer. But, what exactly is the other term?Let's say x1 is an image from the dataset X, then the GroupPooled output from the layer i is the first term. How exactly is the other term here defined? lines found here