torch.max(attention, dim=-1, keepdim=True)[0].expand_as(attention) - attention

[MorrisXu-Driving]

This line may reverse the weight. when MAX - Attention, the positions with max attention weight becomes zero. I also did not find relevant information in the paper. Why add this line?

[nuaawh]

Did you figure out the issue?

[MorrisXu-Driving]

I just deleted this line and do my own experiments with no anomaly spotted.