CenterScaleCrop - Implement inverse method

[jpcenteno80]

Hi, would it be possible to implement an inverse method for the above transform? I would swear that it used to work a few dev versions ago. I just tried to re-run some inference code using this version (MONAI version: 0.5.3+116.gb151b8e) and I run into the following error:

  File "/home/jpcenteno/miniconda3/envs/test_pkgs/lib/python3.9/site-packages/monai/transforms/inverse.py", line 121, in inverse
    raise NotImplementedError(f"Subclass {self.__class__.__name__} must implement this method.")
NotImplementedError: Subclass CenterScaleCropd must implement this method.

Thanks!

[jpcenteno80]

This fixed the issue in some test cases. It is a copy of what was done for the inverse method in the class RandSpatialCropSamplesd:

## implementation of inverse method for CenterScaleCropd
import contextlib
from copy import deepcopy
from math import ceil
from typing import Dict, Hashable, Mapping, Sequence, Union

from monai.config import KeysCollection
from monai.transforms import CenterSpatialCropd
from monai.transforms.inverse import InvertibleTransform
from monai.transforms.transform import MapTransform
from monai.transforms.utils import allow_missing_keys_mode
from monai.utils import ensure_tuple_rep
from monai.utils.enums import InverseKeys

@contextlib.contextmanager
def _nullcontext(x):
    """
    This is just like contextlib.nullcontext but also works in Python 3.6.
    """
    yield x

class CenterScaleCropd(MapTransform, InvertibleTransform):
    """
    Dictionary-based wrapper of :py:class:`monai.transforms.CenterScaleCrop`.
    Args:
        keys: keys of the corresponding items to be transformed.
            See also: monai.transforms.MapTransform
        roi_scale: specifies the expected scale of image size to crop. e.g. [0.3, 0.4, 0.5] or a number for all dims.
            If its components have non-positive values, will use `1.0` instead, which means the input image size.
        allow_missing_keys: don't raise exception if key is missing.
    """

    def __init__(
        self, keys: KeysCollection, roi_scale: Union[Sequence[float], float], allow_missing_keys: bool = False
    ) -> None:
        super().__init__(keys, allow_missing_keys=allow_missing_keys)
        self.roi_scale = roi_scale

    def __call__(self, data: Mapping[Hashable, np.ndarray]) -> Dict[Hashable, np.ndarray]:
        d = dict(data)
        img_size = data[self.keys[0]].shape[1:]
        ndim = len(img_size)
        roi_size = [ceil(r * s) for r, s in zip(ensure_tuple_rep(self.roi_scale, ndim), img_size)]
        self.cropper = CenterSpatialCropd(self.keys, roi_size=roi_size, allow_missing_keys=self.allow_missing_keys)
        cropped = self.cropper(d)
        # CenterSpatialCropd was added to the list. Change to CenterScaleCropd
        for key in self.key_iterator(cropped):
            cropped[str(key) + InverseKeys.KEY_SUFFIX][-1][InverseKeys.CLASS_NAME] = self.__class__.__name__
            cropped[str(key) + InverseKeys.KEY_SUFFIX][-1][InverseKeys.ID] = id(self)        
        return cropped
    
    def inverse(self, data: Mapping[Hashable, np.ndarray]) -> Dict[Hashable, np.ndarray]:
        d = deepcopy(dict(data))
        # We changed the transform name from CenterSpatialCropd to CenterScaleCropd
        # Need to revert that since we're calling CenterSpatialCropd's inverse
        for key in self.key_iterator(d):
            d[key + InverseKeys.KEY_SUFFIX][-1][InverseKeys.CLASS_NAME] = self.cropper.__class__.__name__
            d[key + InverseKeys.KEY_SUFFIX][-1][InverseKeys.ID] = id(self.cropper)
        context_manager = allow_missing_keys_mode if self.allow_missing_keys else _nullcontext
        with context_manager(self.cropper):
            return self.cropper.inverse(d)

[Nic-Ma]

Hi @jpcenteno80 ,

Thanks for raising this up.
Are you using the dev branch or MONAI v0.5.3?
I think we don't have this transform in 0.5.3:
https://github.com/Project-MONAI/MONAI/blob/0.5.3/monai/transforms/croppad/dictionary.py

Thanks.

[jpcenteno80]

Sorry, I did not include the details on the environment. That is true, that transform is not in v0.5.3... It is part of dev(pip install git+https://github.com/Project-MONAI/MONAI#egg=monai):

MONAI version: 0.5.3+116.gb151b8e
Numpy version: 1.20.3
Pytorch version: 1.9.0+cu102
MONAI flags: HAS_EXT = False, USE_COMPILED = False
MONAI rev id: b151b8ebd6c2f7481a4348689c6ac3eee2d5f7ef

Optional dependencies:
Pytorch Ignite version: 0.4.4
Nibabel version: 3.2.1
scikit-image version: 0.18.1
Pillow version: 8.2.0
Tensorboard version: 2.5.0
gdown version: 3.13.0
TorchVision version: 0.10.0+cu102
ITK version: 5.1.2
tqdm version: 4.61.1
lmdb version: 1.2.1
psutil version: 5.8.0

For details about installing the optional dependencies, please visit:
    https://docs.monai.io/en/latest/installation.html#installing-the-recommended-dependencies

[Nic-Ma]

Thanks for your bug, I can reproduce it now.
Let me try to fix it ASAP.

Thanks.

[Nic-Ma]

Submitted PR to fix it: #2383
Thanks.