Label numpy to tensor conversion issue

[relyativist]

Hi,
After composing transform for np.array float32 labels, which have data type and unique:

data = nibabel.load(val_files[0]['seg'])
img_array = data.get_fdata().astype(np.float32)
print(img_array.dtype)
np.unique(img_array)

float32
array([0., 1.], dtype=float32)

I get the output tensor with

After checking output tensor, I get the value with not only 0 and 1:
torch.unique(val_data_example["seg"])

torch.float32
tensor([0.0000, 0.0370, 0.0556, 0.0741, 0.0833, 0.1111, 0.1250, 0.1481, 0.1667,
        0.1852, 0.2222, 0.2500, 0.2593, 0.2778, 0.2963, 0.3333, 0.3704, 0.3750,
        0.3889, 0.4074, 0.4167, 0.4444, 0.4815, 0.5000, 0.5185, 0.5556, 0.5833,
        0.5926, 0.6111, 0.6296, 0.6667, 0.7037, 0.7222, 0.7407, 0.7500, 0.7778,
        0.8148, 0.8333, 0.8519, 0.8750, 0.8889, 0.9167, 0.9259, 0.9444, 0.9630,
        1.0000])

I suppose, behaviour of translating from numpy to tensor is wrong. Are there any monai transformations that could resolve this?

[relyativist]

Load dataset:

check_dataset = Dataset(data=train_files, transform=train_trans)
val_data_example = check_dataset[5]

Transformations:

train_trans = Compose(
    [
        LoadImaged(keys=["image", "seg"]),
        EnsureChannelFirstd(keys=["image", "seg"]),
        Resized(keys=["image", "seg"], spatial_size=(128,128,128)),
        ScaleIntensityd(keys=["image"], minv=0.0, maxv=1.0, channel_wise=True),
        EnsureTyped(keys=["image", "seg"]),
    ]
)
val_transf = Compose(
    [
        LoadImaged(keys=["image", "seg"]),
        EnsureChannelFirstd(keys=["image", "seg"]),
        Resized(keys=["image", "seg"], spatial_size=(128,128,128)),
        ScaleIntensityd(keys=["image"], minv=0.0, maxv=1.0, channel_wise=True),
        EnsureTyped(keys=["image", "seg"]),
    ]
)

[Nic-Ma]

Please set different interpolation modes for image and label:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/spatial/dictionary.py#L809

Thanks.