The code and detailed implementation of Figure 4 and Figure 5 in the paper mPLUG-Owl2

[Zlatan-Ibrahi]

I would like to analyze the attention map of my own trained model, but I am not very clear about some details. For example, do we take the average of the attention maps across multiple heads? Could you provide the code for this?

[GasolSun36]

same question, any solutions?