GitHub - leffff/pytorch-hypersphere (original) (raw)

pytorch-hypersphere

A simple yet efficient implementation of transformation from Euclidian to Hyperspherical (N-spherical) space

The library uses this Wikipedia article as a basis

Package listing:

• layers

import torch

from pytorch_hypersphere.layers import ToHyperSphere, ToEuclidean

ths = ToHyperSphere(16) # initialize transformation layer

te = ToEuclidean(16) # initialize transformation layer

x_eucl = torch.randn((4, 16)) # random floats in euclidian space

x_sphere = ths(x_eucl) # transformation to hyperspherical

x_eucl_2 = te(x_sphere) # transformation back to euclidean

• functional

import torch

from pytorch_hypersphere.functional import to_hypersphere, to_euclidean

x_eucl = torch.randn((4, 16)) # random floats in euclidian space

x_sphere = to_hypersphere(x_eucl) # transformation to hyperspherical

x_eucl_2 = to_euclidean(x_sphere) # transformation back to euclidean

• rand

from pytorch_hypersphere.random import euclidean_randn_spherical, nsphere_randn_spherical

random_points_on_sphere_in_euclidean = euclidean_randn_spherical(shape=(4, 16), stretch_coefficient=2) # generate points randomly distributed on a sphere, in euclidean coordinates, with radius of 2 random_points_on_sphere_in_nsphere = nsphere_randn_spherical(shape=(4, 16), stretch_coefficient=1) # generate points randomly distributed on a sphere, in spherical coordinates, with radius of 1