search

keras-team / keras

Deep Learning for humans
http://keras.io/
Apache License 2.0
61.26k stars 19.38k forks source link

Conv1D on custom shaped data #19900

Open Wasim04 opened 1 week ago

Wasim04 commented 1 week ago

Hi,

Just wondering if there is a way to use Conv1D in a heapix data. That is, it can handle, resolution(nside) and pixel orientation (nested or ring).

Many Thanks Wasim

mehtamansi29 commented 1 week ago

Hi @Wasim04 -

Healpix stands for Hierarchical Equal Area isoLatitude Pixelization of a sphere where partitioning sphere at higher resolutions. So this partition of sphere surface has different size pixels and which is represented as 2D array. So it is better to use Conv1D and Conv2D both are not suitable for healpix image data. Conv1D can create convolution on single dimension data while Conv2D can create convolution on two dimensional data.

JAX library is better option to implement spherical CNNs.

Definitely we can do experiment with Conv1D with healpix data(by projecting MNIST dataset onto HEALPix), but it will not generate good result for large spherical data.

Let me know if you required more details. Thanks..!!