Visualization of attention regions

[tong8080]

Hello, how do you draw the figure 8(Visualization of attention regions) in the paper? Is it reflected in the code?

[husthuaan]

@tong8080 Sorry that it's not reflected in the code. But it's not difficult to implement: during the decoding process, you can obtain the attention weights for all image regions and use alpha blending to draw the figure.

[tong8080]

Thanks, I will have a try.

[2014gaokao]

@tong8080 have you implement it? pretty difficult for me.

[invisprints]

Same problem, has anybody solved it?

[nilinykh]

Hello everyone, same question here: how do I do it during the decoding phase? During decoding, we have multiple attention weights, since there are several heads and also several layers, right? What would be the way to combine them in such fashion, that something similar to Fig. 8 can be created? Or am I misunderstanding something?

[gabrielsantosrv]

Hello everyone, same question here: how do I do it during the decoding phase? During decoding, we have multiple attention weights, since there are several heads and also several layers, right? What would be the way to combine them in such fashion, that something similar to Fig. 8 can be created? Or am I misunderstanding something?

I'm also confused about the same points