BigGAN: Class Conditional Batch Norm 2D

Course Q&A Generative Adversarial Networks (GANS) Build Better Generative Adversarial Networks
1

Hey guys,
I have 2 small queries with the implementation of the Class Conditional Batch Norm 2D layer in the BigGAN. I have attached the apt screenshots for reference below.

2
21603×757 54.6 KB

In the explanation for the same, it has been simply said that gamma and beta are obtained via transformations of the class embedding c, however, in the code, Spectral Norm has been used over the linear transformation. Can anyone tell me what is the role of Spectral Norm here?

1
11495×593 99.1 KB

Additionally, when we obtained class_scale, we have added 1 to the output of the transformed outputs, which is nowhere mentioned in the theory. Can anyone let me know about the significance of this operation?

Thanks in advance!

2

Hey @mentor,
Can you help me with my query please?

3

Hi @Elemento, Thank you for the question. Our mentors will answer this shortly.
@sohonjit.ghosh, @shreyasvedpathak, @pedrorohde I would like to request you to take some time and answer this query.

Thank you,
Sharob