How to realize the attention maps？

[autonomousvision]

How to realize the attention maps？

[ap229997]

1. Transformer has multiple attention heads and attention layers. In our visualizations, we consider the final attention layer from each of the 4 attention heads. So, even though the query token is the same, different heads can attend to different regions in the input.

2. I have not explicitly visualized cases involving traffic light change so I don't have a good answer but that is an interesting point.

3. I agree that it is hard to reason about some of these mappings. In general, we observed that the transformer can learn associations among traffic lights with vehicles but the association is not always correct (eg. with the wrong direction of traffic). In this particular case, I think different sides of the intersection are mapped to each other (I've also observed this to be true in some other visualizations as well).

[Watson52]

Thank you for your quick response. But I still don't quite understand about the third point that sides of the intersection are mapped to each. Could you please give me some other examples about the visualizations?

[ap229997]

It is just my hypothesis and I am not completely sure about it. You can look at the 2nd visualization in the last row where vehicles on the left seem to be mapped to a region around the traffic light on the right side of the intersection.

I agree that this argument is not entirely convincing but I don't have a better answer right now.

[Watson52]

I see. I will try to run the code later and ask you when I meet a problem. Thank you! Btw, is there another way to download the dataset? I download it slowly on Amazon or sometimes I can't connect to it.

[ap229997]

Unfortunately, there is no other way to download the dataset.

[Watson52]

I see. Thank you :)