Scipy and Pytorch backends give different results at order=3 #2
Description
When providing order=3 to each backend, the results are wildly different. Here's code:
import numpy as np
import resize.scipy as resize_sp
import resize.pytorch as resize_pt
import torch
x1 = torch.arange(5).float()[:, None]
x2 = torch.arange(7).float()[None, :]
x_pt = (x1 + x2).float()
x_sp = x_pt.numpy()
# Batch, channel for pytorch
x_pt = x_pt[None, None, ...]
dxyz = (0.7, 2.4)
for order in [0, 1, 3]:
y_pt = resize_pt.resize(x_pt, dxyz, order=order)
y_sp = resize_sp.resize(x_sp, dxyz, order=order)
print(f"{order=}, Close? {np.allclose(y_pt.numpy(), y_sp)}")
When order=3, the final array values are:
tensor([[[[0.4432, 2.9393, 5.4352],
[1.0080, 3.5040, 6.0000],
[1.8460, 4.3420, 6.8380],
[2.5040, 5.0000, 7.4960],
[3.1620, 5.6580, 8.1540],
[4.0000, 6.4960, 8.9920],
[4.5647, 7.0608, 9.5567]]]])
vs
array([[0.49109966, 2.9585092 , 5.4259186 ],
[1.0636655 , 3.531075 , 5.9984846 ],
[1.8582219 , 4.3256316 , 6.793041 ],
[2.5325904 , 5. , 7.4674096 ],
[3.206959 , 5.6743684 , 8.141778 ],
[4.0015154 , 6.468925 , 8.936335 ],
[4.5740814 , 7.0414906 , 9.508901 ]], dtype=float32)