The margin and distance type in your implementation

Thanks for your greate job! I found that the hyper-parameter ‘margin’ and the ‘distance type’ in your code is the opposite of your paper report： In your code:(detectron2/modeling/roi_heads/fast_rcnn.py/FastRCNNOutputLayers) self.hingeloss = nn.HingeEmbeddingLoss(2) means the margin is 2.

distances = torch.cdist(fg_features, torch.stack(all_means).cuda(), p=self.margin) where in your code the p(self.margin) is 10, which means a 10-Norm

But in your papar: You said that the distance is Euclidean distance so the p is should be 2 and the margin should be 10.

I suspect you are writing the reverse Looking forward to your apply.

[JosephKJ]

Its a writing and variable name issue. It will not affect the results.

[1170300714]

收到，感谢您的来件