Visualizing Attention for Sequence Data in the Frequency domain

I have trouble finding out how i can map my attention weights onto my input data.

My input data is in a 512*1 Cell, it represents a vibration in the frequency-domain, after going through different layers (mainly 2 CNN and 2 Bi-GRU), i have a high level representation of it(480*1), that enters a self attention layer.

I am able to extract the attention weights (256*1), but I am unsure on how to map them to my input sequence, am i able to just upscale the attention vector and overlay it onto my input?

I've added a screenshots of the model structure, to help you understand the problem.

The whole problem is further illustrated in this research paper.

I would be grateful for anyone, who can help me understand how I need to proceed.

Answer 1

[Kaustab Pal]

Hi @Luca,

To upscale your attention weights from a 256x1 dimension to 512x1, you can utilize the "imresize" function.

You can refer to the ‘imresize’ function documentation available here: https://www.mathworks.com/help/matlab/ref/imresize.html

Below is an example demonstrating how to achieve this:

A = randn(256,1); % dummy data of size 256x1 

B = imresize(A,[512,1]); % resize to required size 

size(B) % return an image B of size 512x1 

ans = 1x2

512 1

<mw-icon class=""></mw-icon>

<mw-icon class=""></mw-icon>

I hope this solves your issue.