Are radius and angles trainable?

[huzi96]

First of all, thank you for sharing the code!
I read through the OpenReivew posts and I noticed that only the network weights should be optimized. All other parameters such as the initialization schema of the basis encoding function are considered hyper-parameters. However, I see in your code

PNF/models/sbn_2dfan.py

Line 20 in 22b6ce9

self.radius = nn.Parameter(

that the angles and radius are configured as trainable parameters. Will them be optimized during training? If so, how can you guarantee band limits?

[stevenygd]

You should fix them in order to get precise constrained subbands. But training the radius and angel together usually leads to slightly better performance.