search

karpathy / convnetjs

Deep Learning in Javascript. Train Convolutional Neural Networks (or ordinary ones) in your browser.
MIT License
10.9k stars 2.04k forks source link

[Suggestion for "Painting"] Use of Fourier Features? #121

Open UPstartDeveloper opened 2 years ago

UPstartDeveloper commented 2 years ago

Hello,

On the page titled "Image Painting" I wanted to suggest passing the (x, y) points through a Fourier feature mapping. The details on this are described here in this 2020 NeurIPS paper by Google Research (including example code).

Admittedly I'm not sure if this would be out of scope for an intro example for this library, but it would be very cool to see the neural net be able to paint the cat clearly.

Btw thanks for sharing this great work! I originally found this repo when I was in undergrad, it was part of what got me excited about computer vision :)

371148606 commented 2 years ago

您好，我是薛文毅，已收到您的邮件，一会查收，谢谢！！  

UPstartDeveloper commented 2 years ago

Specifically for implementation, it might be good to find a NumPy equivalent for JS. Reading around it looks like numjs is a good option

Update: could also be a good idea to write a new type of layer class for this library, for easier incorporation into linear models

UPstartDeveloper commented 1 year ago

^hi again @371148606 - I figured out a way to implement this just using the existing utility functions (credit @karpathy :) ). Would be great to get some feedback!