1 Dimension Only?

[jeongjoonpark]

Hi I wonder if this code works for 2 or 3 dimensional data!

Thanks.

[patrick-kidger]

Depends what you mean!

The first example in the README considers three different dimensions (=channels) at the same time:

import torch
from torchcubicspline import(natural_cubic_spline_coeffs, 
                             NaturalCubicSpline)

length, channels = 7, 3
t = torch.linspace(0, 1, length)
x = torch.rand(length, channels)
coeffs = natural_cubic_spline_coeffs(t, x)
spline = NaturalCubicSpline(coeffs)
point = torch.tensor(0.4)
out = spline.evaluate(point)

If you mean that you want to have multiple "lengths", then I'm afraid not. (Good ways of doing splines isn't even a solved problem in that context.)

Does that answer your question?

[yanconglin]

Hi, Patrick Kidger,

Thanks for sharing!
I have a similar question: interpolating on non-regular 2D/3D data.
Seems like your approach only interpolates over the time dimesion (1-D data).
Now I have a 2D non-regular grid, and want to upsample over this grid.
Since the Pytorch interploation only works with regular grids (e.g. images), I am looking for other implementations.

x = np.linspace(-3, 3, 100) # non regular grids
y = np.linspace(-5, 5, 100)
z = np.linspace(-9, 9, 100)

X,Y, Z = np.meshgrid(x,y,z,indexing='ij')
values = np.randn((100,))

interp = custom_Interp3D_algorithm((x,y,z), values)

// sample points
new_x = np.linspace(-3, 3, 400)
new_y = np.linspace(-5, 5, 400)
new_z = np.linspace(-9, 9, 400)
// output new values: [400, ]
new_values = interp((new_x, new_y, new_z))

Once can also wrap this into BXCXHXW tensor, where HX W is the 2d grid.
Was wondering if you have such an implemtation.
Thanks a lot!

Yancong

[patrick-kidger]

Hi @yanconglin - I'm afraid not, is the short answer.

Producing higher-dimensional interpolants, on irregular data, is much harder than the 1D case. Methods for doing so are much less standard than in the 1D case.