InterpolationMode in PyTorch

Buy Me a Coffee☕

*Memos:

• My post explains Resize().
• My post explains RandomRotation().
• My post explains OxfordIIITPet().

You can set InterpolationMode as shown below:

*Memos:

• In PyTorch to resample(resize) data, there are 9 interpolation modes Nearest(Nearest-neighbor) which matches the OpenCV’s INTER_NEAREST which is buggy, Nearest-exact(Nearest-neighbor) which matches the Scikit-Image's and PIL's(Pillow's) nearest neighbour which isn't buggy, Linear, Bilinear, Trilinear, Bicubic, Box, Hamming and Area according to Source code for torchvision.transforms.functional and torch.nn.functional.interpolate. *I will say again that Nearest is buggy while Nearest-exact isn't buggy.
• interpolation argument with antialias argument is for Resize(), RandomResizedCrop(), etc.
• interpolation argument without antialias argument is for RandomRotation(), RandomAffine(), RandomPerspective(), ElasticTransform(), RandAugment(), TrivialAugmentWide(), AutoAugment(), AugMix(), etc.

from torchvision.datasets import OxfordIIITPet
from torchvision.transforms.v2 import Resize, RandomRotation
from torchvision.transforms.functional import InterpolationMode

origin_data = OxfordIIITPet(
    root="data",
    transform=None
)

import matplotlib.pyplot as plt

def show_rimages(im, s=None, ip=None):
    title = "s" + str(s) + "ip" + str(ip).split(".")[1] + "a"
    title1 = title + "True_data"
    title2 = title + "False_data"
    plt.figure(figsize=[10, 8])
    for i in range(1, 3):
        plt.subplot(1, 2, i)
        r = Resize(size=s, interpolation=ip,
                   antialias=True if i == 1 else False)
        plt.title(label=title1 if i == 1 else title2, y=1, fontsize=14)
        plt.imshow(X=r(im))
    plt.tight_layout()
    plt.show()

def show_rrimages(im, d=None, ip=None):
    plt.figure(figsize=[10, 8])
    for i in range(2):
        if isinstance(d, collections.abc.Sequence):
            d1 = str(d[0]) if d[0] >= 0 else "n" + str(-1*d[0])
            d2 = str(d[1]) if d[1] >= 0 else "n" + str(-1*d[1])
            dpart = d1 + d2 if "n" in d2 else d1 + "_" + d2
        else:
            dpart = str(d)
        title = "d" + dpart + "ip" + str(ip[i]).split(".")[1] + "_data"
        plt.subplot(1, 2, (i+1))
        rr = RandomRotation(degrees=d, interpolation=ip[i])
        plt.title(label=title, y=1, fontsize=14)
        plt.imshow(X=rr(im))
    plt.tight_layout()
    plt.show()

plt.figure(figsize=[7, 9])
plt.title(label="s500_394origin_data", fontsize=14)
plt.imshow(X=origin_data[0][0])
plt.show()
print()
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.NEAREST)
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.NEAREST_EXACT)
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.BILINEAR)
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.BICUBIC)
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.BOX)
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.HAMMING)
show_rimages(im=origin_data[0][0], s=50, ip=InterpolationMode.LANCZOS)
print()
show_rrimages(im=origin_data[0][0], d=[45, 45],
              ip=[InterpolationMode.NEAREST, InterpolationMode.NEAREST_EXACT])
show_rrimages(im=origin_data[0][0], d=[45, 45],
              ip=[InterpolationMode.BILINEAR, InterpolationMode.BICUBIC])

---

---