About the distance metric

[XiaLiPKU]

CFBI/networks/layers/matching.py

Line 38 in 5da872e

d = xs + ys - 2. * torch.matmul(x, torch.t(y))

The distance metric is simply the l2-distance, as shown above.
However, in your paper, the distance is defined as Eq.1.
So, which one is better?

[z-x-yang]

This is just the same distance in our paper.

||x-y||^2=sum(x^2)-2*sum(xy)+sum(y^2)

The latter format is more computationally efficient.

[XiaLiPKU]

This is just the same distance in our paper.

||x-y||^2=sum(x^2)-2*sum(xy)+sum(y^2)

The latter format is more computationally efficient.

But in the paper, the equation is 1 - 2 / (1 + exp(||x -y||^2 + bias)

[XiaLiPKU]

But in the paper, the equation is 1 - 2 / (1 + exp(||x -y||^2 + bias)
This difference doesn't influence the minimum indexes but influences the minimum values

[z-x-yang]

1 - 2 / (1 + exp(||x -y||^2 + bias).

See here.

[XiaLiPKU]

1 - 2 / (1 + exp(||x -y||^2 + bias).

See here.

Got it, Thank you very much!

[z-x-yang]

Sure thing.