GAP sine normalization via input layers

[jdeschamps]

Problem

GAP includes normalization in the first few layers using sine transforms. As opposed to statistics normalization, this removes the need to do a first-pass on the data before training, and is compatible (theoretically) with images with diverse dynamical ranges.

References

• In the code
• Paper