features being extracted by the continuous information in the infoGAN

[favrei]

Hello TAs,

In the reference code, and in the most of the resource from the internet, 2 dimensions of continuous features are separate from the input noise. I am wondering why are font width and the angles always being extracted in this configuration, independent to the hyperparameters and kept consistent among different runs.

Thank You,

[chunchih]

The generator's input is divided into "noise", "dis_c"(class info), "continuous"(some continuous info), so I guess most obvious part and continuous data are angle and font-width. But we still can't guarantee what kind of feature the network detects.