Lazy Transform Doesn't Apply Defined Interpolation Method

[lukaskrl]

When using lazy transform, specifying an interpolation method still applies the default bilinear interpolation to the image. For instance, when resizing an image mask with nearest neighbors interpolation, the mode is reset to None (defaulting to bilinear), leading to an incorrect mask.

from monai.transforms import (
    Compose,
    Resized
)
import numpy as np

def generateTransform(image_size = (15, 15, 15)):
    return Compose(
        [
            Resized(
                keys=["image", "label"],
                spatial_size=image_size,
                mode=("bilinear", "nearest"),
            ),
        ],
        lazy=True,
    )

image = np.random.rand(10, 10, 10)
mask = np.random.randint(0, 2, size=(10, 10, 10))

print(np.unique(mask))
#[0 1]

data_dict = {'image': image, 'label': mask}
sample = generateTransform()(data_dict)

print(np.unique(sample['label']))
#[0.         0.02777778 0.08333334 0.1388889  0.16666667 0.25 0.2777778  0.30555555 0.41666666 0.5 ...]

Expected behavior
The interpolation method applied to the mask should be nearest neighbours. The method defaults to bilinear, giving us a mask with float values. It seems the interpolation mode is not passed through the composed transforms when using the lazy flag.

Environment
Python 3.10.12.
monai 1.3
numpy 1.24.4

[KumoLiu]

Hi @lukaskrl, thanks for the reporting.
You should add overrides in Compose.

trans = Compose([...], overrides={"image": {"mode": "bilinear"}, "label": {"mode": "nearest"}})

Hope it helps, thanks.

[KumoLiu]

For the dtype one which is a known issue, as reported here.

[KumoLiu]

Convert to the discussion now, feel free to recreate one if face any issues.

[atbenmurray]

@KumoLiu I'd like to just double check that there isn't another issue. The overrides approach does avoid the problem, but there may still be an underlying problem.

[KumoLiu]

Hi @atbenmurray, Which one are you referring to? And what is the 'underlying problem', can you elaborate on that in more detail? Thanks.