Attribute Re-Weighting Module

[zarakikun]

Is there attribute re-weighting module in the training net?

[Yukun-Huang]

No, I did not introduce such a module into the network.

[jpainam]

Hi. I tried to introduce the re-weighting module. But can't figure out the right values. According to the original paper

m = 28 # the number of attributes
\tilde{a} \in R^{1x28}

After the element-wise multiplication

a = c . a  # R^{mx1} . R^{1xm} 
a \in R^{m x m}

A global feature representation is [1, 2048]. The paper says, a \in ([28, 28]) concatenated with the global features ([1, 2048]). Anyone knows how to concatenate these two features vectors.

[28, 28] x [1, 2048]

[abhigoku10]

@jpainam where you able to implement this functionality ??