dual branch and output score normalization

[KarenAssaraf]

Hey!
Thanks for beeing so responsive.
I have another question again...

in the dual branch,
FC score is a fully connect that ends with a RELU and fc_weight branch ends with a sigmoid. Am I correct that it makes score q between 0 and + infinity.

Is the network supposed to output normalized scores?

[TianheWu]

Hi! It is a good question! In different perceptual scale, it has different quality score. For easily training the network, we normalize the all dataset score to 0-1. The output should between 0-1. If you want to obtain the original score, you can wrtie a reversed function of normalization.

[KarenAssaraf]

Thanks a lot!
I am trying to use your network as a backbone, using koniq - pretrained.
I am doing a pairwise finetuning (siamese network scale) . where my data is paired images, and using ranking margin loss.
The think is I see the network output values between -5 and 5... so I was wondering how could be.

[KarenAssaraf]

Hi, actually, from the code of the dual branch, even though we normalize the label score, it could be that the output is greater than 1.
I think this is due to fc_score branch, that ends with a RELU meaning fc_score value could be greater than 1.
if fc_score is greater than 1, then the dual branch output value _s = torch.sum(f * w) / torch.sum(w), would also be greater than 1.
Am I correct?