convnetjs
convnetjs copied to clipboard
Fourier features
Closes #121. It's not perfect but I'd love to get folks' feedback on the current state of this code on this branch!
Change Summary:
- I added a new layer class called
FourierFeatureLayer
, and demo its usage by including it in the default neural network defined in theimage_regression.js
file - I added my contact link in
image_regression.html
(in addition to keeping @karpathy 's Twitter handle link). Happy to keep working on this example and/or engage in further discussion about it!
Known limitations:
- For some reason, passing a value greater than
0
forgaussian_mapping_scale
causes the loss of the model to plateau. For now I've set it so that it defaults to use using sinusoidal functions in the mapping (please the paper I'm working off of in case that sentence doesn't make sense). - the addition of the Fourier Feature layer in
image_regression.js
causes the wall clock time for training to increase, and increases proportional to how highgaussian_mapping_scale
.
您好,我是薛文毅,已收到您的邮件,一会查收,谢谢!!