Setting InterpolationMode.NEAREST_EXACT to interpolation argument of RandomRotation() works against the error message

[hyperkai]

🐛 Describe the bug

Setting InterpolationMode.BOX to interpolation argument of RandomRotation() gets the error message as shown below:

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

rr = RandomRotation(degrees=[90.0, 90.0], interpolation=InterpolationMode.BOX)
                                        # ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ 
import matplotlib.pyplot as plt

plt.imshow(X=rr(my_data[0][0])) # Error
plt.show()

ValueError: Image.Resampling.BOX (4) cannot be used. Use Image.Resampling.NEAREST (0), Image.Resampling.BILINEAR (2) or Image.Resampling.BICUBIC (3)

But setting InterpolationMode.NEAREST_EXACT to interpolation argument of RandomRotation() works against the error message as shown below:

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

rr = RandomRotation(degrees=[90.0, 90.0], interpolation=InterpolationMode.NEAREST_EXACT)
                                        # ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ 
import matplotlib.pyplot as plt

plt.imshow(X=rr(my_data[0][0]))
plt.show()

So, the error message should also have Image.Resampling.NEAREST_EXACT (1) as shown below:

ValueError: Image.Resampling.BOX (4) cannot be used. Use Image.Resampling.NEAREST (0), Image.Resampling.NEAREST_EXACT (1), Image.Resampling.BILINEAR (2) or Image.Resampling.BICUBIC (3)

Versions

import torchvision

torchvision.__version__ # '0.20.1+cpu'

[NicolasHug]

Hi @hyperkai

The error message you're seeing comes from PIL, not from torchvision. You passed a PIL image and for PIL images, NEAREST is equivalent to NEAREST_EXACT (they will be slightly different for tensors).