Problem when dim x not same as dim y

[monabf]

Hi,

Thanks a lot for the implementation. However, I have a problem with the dimensions. For example, x is a 1D grid of 100 points, y is a 2D grid of 100 points, and I want to interpolate the 2D value of y at one point in x denoted xnew. Then I should obtain a 2D value for ynew, which should be the interpolation of each axis of y at the interpolated coordinate xnew. However, I obtain a scalar ynew. Conversely, when x is a 2D grid and y is a 1D grid, I obtain ynew of dimension 2 for xnew of dimension 2. Am I missing something here?

Below two small working examples, with respectively

• x (1,N), y (2,N), xnew (1,P), which should give ynew (2,P)
• x (N,), y (2,N), xnew (P,), which should also give ynew (2,P)

Thanks a lot for your help!

import torch
import matplotlib.pyplot as plt
import time
import numpy as np
from torchinterp1d import Interp1d


if __name__ == "__main__":
    # problem dimensions
    Dx = 1
    Dy = 2
    N = 10
    P = 1

    yq_gpu = None
    yq_cpu = None
    x = torch.linspace(0, 10, Dx*N).view(Dx,N)
    y = torch.linspace(0, 10, Dy*N).view(Dy, -1)
    xnew = torch.rand(Dx, P) * 10

    # calling the cpu version
    t0_cpu = time.time()
    yq_cpu = Interp1d()(x, y, xnew, yq_cpu)
    t1_cpu = time.time()

    print(x.shape, y.shape, xnew.shape, yq_cpu.shape)
    print(x, y, xnew, yq_cpu)



    # problem dimensions
    Dy = 2
    N = 10
    P = 1

    yq_gpu = None
    yq_cpu = None
    x = torch.linspace(0, 10, N).view(N,)
    y = torch.linspace(0, 10, Dy*N).view(Dy, -1)
    xnew = torch.rand(P) * 10

    # calling the cpu version
    t0_cpu = time.time()
    yq_cpu = Interp1d()(x, y, xnew, yq_cpu)
    t1_cpu = time.time()

    print(x.shape, y.shape, xnew.shape, yq_cpu.shape)
    print(x, y, xnew, yq_cpu)

Output:

torch.Size([1, 10]) torch.Size([2, 10]) torch.Size([1, 1]) torch.Size([1, 1])
tensor([[ 0.0000,  1.1111,  2.2222,  3.3333,  4.4444,  5.5556,  6.6667,  7.7778,
          8.8889, 10.0000]]) tensor([[ 0.0000,  0.5263,  1.0526,  1.5789,  2.1053,  2.6316,  3.1579,  3.6842,
          4.2105,  4.7368],
        [ 5.2632,  5.7895,  6.3158,  6.8421,  7.3684,  7.8947,  8.4211,  8.9474,
          9.4737, 10.0000]]) tensor([[5.6551]]) tensor([[2.6787]])
torch.Size([10]) torch.Size([2, 10]) torch.Size([1]) torch.Size([1, 1])
tensor([ 0.0000,  1.1111,  2.2222,  3.3333,  4.4444,  5.5556,  6.6667,  7.7778,
         8.8889, 10.0000]) tensor([[ 0.0000,  0.5263,  1.0526,  1.5789,  2.1053,  2.6316,  3.1579,  3.6842,
          4.2105,  4.7368],
        [ 5.2632,  5.7895,  6.3158,  6.8421,  7.3684,  7.8947,  8.4211,  8.9474,
          9.4737, 10.0000]]) tensor([0.9259]) tensor([[0.4386]])

[aliutkus]

dear @monabf, maybe you should just duplicate the xnew variable ?

you're right, the doc from the readme and the docstring from the function don't seem to match. As it is implemented now, it looks like out is same shape as xnew. I may have a look into this

[monabf]

Hi @aliutkus, you're right, using x = x.expand(Dy,-1) seems to solve the problem and avoids slicing away part of the y. Thanks for looking into it! It would be great to have this corrected or documented indeed

[zfhxi]

To explore which of x and xnew should be expanded, I write a demo:

import torch
from utils.gpu.torch_interp1d import interp1d
import scipy.interpolate as syinterpolate

if __name__ == "__main__":
    N = 100
    D = 1024
    P = 30
    x = torch.arange(N).view(-1, N)  # -> 1,N
    y = torch.randn([D, N])
    xnew = torch.linspace(0, N, P)
    ynew = interp1d(x, y, xnew)
    ynewv2 = interp1d(x.expand(D, -1), y, xnew)
    ynewv3 = interp1d(x, y, xnew.expand(D, -1))
    ynewv4 = interp1d(x.expand(D, -1), y, xnew.expand(D, -1))

    # scipy interp1d
    x = x.squeeze().cpu().numpy()  # N
    y = y.transpose(-1, -2).cpu().numpy()  # N,D
    xnew = xnew.cpu().numpy()  # P
    fit_func = syinterpolate.interp1d(x, y, kind="linear", axis=0, fill_value="extrapolate")
    ynewv5 = fit_func(xnew)
    ynewv5 = torch.from_numpy(ynewv5).float().transpose(-1, -2)  # D,P

    print(ynew.shape)  # 1,P
    print(ynewv2.shape)  # D,P
    print(ynewv3.shape)  # D,P
    print(ynewv4.shape)  # D,P
    # print((ynewv2 == ynewv3).all())  # False
    # print((ynewv2 == ynewv4).all())  # True
    # print((ynewv2 == ynewv5).all())  # may be False, caused by computational accuracy of the cpu

    print((ynewv2 - ynewv3).abs().sum())  # >0
    print((ynewv2 - ynewv4).abs().sum())  # 0
    print((ynewv2 - ynewv5).abs().sum())  # slightly close to 0
    print((ynewv3 - ynewv4).abs().sum())  # >0
    print((ynewv3 - ynewv5).abs().sum())  # >0
    print((ynewv4 - ynewv5).abs().sum())  # slightly close to 0

    # conclusion: ynewv2 is equal to ynewv4 and they both are close to ynewv5

So, we should expand x.