torchio.transforms.transform

Complexity

Total Complexity

89

Size/Duplication

Total Lines	421
Duplicated Lines	0 %

Importance

Changes

0

25 Methods

Rating	Name	Duplication	Size	Complexity
A	Transform.__init__()	0	9	1
A	Transform.__call__()	0	30	5
A	DataToSubject._get_subject_from_tensor()	0	3	1
A	DataToSubject._get_subject_from_sitk_image()	0	4	1
F	Transform.parse_range()	0	104	21
A	DataToSubject._parse_subject()	0	8	2
A	Transform.sitk_to_nib()	0	3	1
A	DataToSubject._get_subject_from_image()	0	3	1
A	Transform.to_range()	0	6	2
A	Transform.add_transform_to_subject_history()	0	11	2
A	Transform.parse_interpolation()	0	13	3
A	Transform.parse_probability()	0	10	3
C	Transform.parse_params()	0	18	10
A	DataToSubject.__init__()	0	14	1
A	Transform.__repr__()	0	10	4
A	Transform.get_arguments()	0	6	1
C	DataToSubject.get_output()	0	26	11
A	DataToSubject._parse_tensor()	0	8	2
A	Transform.nib_to_sitk()	0	3	1
A	DataToSubject._get_subject_from_dict()	0	11	3
A	Transform.inverse()	0	6	2
A	Transform.name()	0	3	1
B	DataToSubject.get_subject()	0	31	8
A	Transform.is_invertible()	0	2	1
A	Transform.apply_transform()	0	3	1

import copy
import numbers
from abc import ABC, abstractmethod
from typing import Optional, Union, Tuple, List

import torch
import numpy as np
import nibabel as nib
import SimpleITK as sitk

from .. import TypeData, DATA, AFFINE, TypeNumber
from ..data.subject import Subject
from ..data.image import Image, ScalarImage
from ..utils import nib_to_sitk, sitk_to_nib, to_tuple
from .interpolation import Interpolation


TypeTransformInput = Union[
    Subject,
    Image,
    torch.Tensor,
    np.ndarray,
    sitk.Image,
    dict,
    nib.Nifti1Image,
]


class Transform(ABC):
    """Abstract class for all TorchIO transforms.

    All subclasses should overwrite
    :py:meth:`torchio.tranforms.Transform.apply_transform`,
    which takes data, applies some transformation and returns the result.

    The input can be an instance of
    :py:class:`torchio.Subject`,
    :py:class:`torchio.Image`,
    :py:class:`numpy.ndarray`,
    :py:class:`torch.Tensor`,
    :py:class:`SimpleITK.image`,
    or a Python dictionary.

    Args:
        p: Probability that this transform will be applied.
        copy: Make a shallow copy of the input before applying the transform.
        keys: Mandatory if the input is a Python dictionary. The transform will
            be applied only to the data in each key.
    """
    def __init__(
            self,
            p: float = 1,
            copy: bool = True,
            keys: Optional[List[str]] = None,
            ):
        self.probability = self.parse_probability(p)
        self.copy = copy
        self.keys = keys

    def __call__(
            self,
            data: TypeTransformInput,
            ) -> TypeTransformInput:
        """Transform data and return a result of the same type.

        Args:
            data: Instance of 1) :py:class:`~torchio.Subject`, 4D
                :py:class:`torch.Tensor` or NumPy array with dimensions
                :math:`(C, W, H, D)`, where :math:`C` is the number of channels
                and :math:`W, H, D` are the spatial dimensions. If the input is
                a tensor, the affine matrix will be set to identity. Other
                valid input types are a SimpleITK image, a
                :py:class:`torch.Image`, a NiBabel Nifti1 Image or a Python
                dictionary. The output type is the same as te input type.
        """
        if torch.rand(1).item() > self.probability:
            return data
        data_parser = DataToSubject(data, keys=self.keys)
        subject = data_parser.get_subject()
        if self.copy:
            subject = copy.copy(subject)
        with np.errstate(all='raise'):
            transformed = self.apply_transform(subject)
        self.add_transform_to_subject_history(transformed)
        for image in transformed.get_images(intensity_only=False):
            ndim = image[DATA].ndim
            assert ndim == 4, f'Output of {self.name} is {ndim}D'
        output = data_parser.get_output(transformed)
        return output

    def __repr__(self):
        if hasattr(self, 'args_names'):
            names = self.args_names
            args_strings = [f'{arg}={getattr(self, arg)}' for arg in names]
            if hasattr(self, 'invert_transform') and self.invert_transform:
                args_strings.append('invert=True')
            args_string = ', '.join(args_strings)
            return f'{self.name}({args_string})'
        else:
            return super().__repr__()

    @abstractmethod
    def apply_transform(self, subject: Subject):
        raise NotImplementedError

    def add_transform_to_subject_history(self, subject):
        from .augmentation import RandomTransform
        from . import Compose, OneOf, CropOrPad
        call_others = (
            RandomTransform,
            Compose,
            OneOf,
            CropOrPad,
        )
        if not isinstance(self, call_others):
            subject.add_transform(self, self.get_arguments())

    @staticmethod
    def to_range(n, around):
        if around is None:
            return 0, n
        else:
            return around - n, around + n

    def parse_params(self, params, around, name, make_ranges=True, **kwargs):
        params = to_tuple(params)
        if len(params) == 1 or (len(params) == 2 and make_ranges):  # d or (a, b)
            params *= 3  # (d, d, d) or (a, b, a, b, a, b)
        if len(params) == 3 and make_ranges:  # (a, b, c)
            items = [self.to_range(n, around) for n in params]
            # (-a, a, -b, b, -c, c) or (1-a, 1+a, 1-b, 1+b, 1-c, 1+c)
            params = [n for prange in items for n in prange]
        if make_ranges and len(params) != 6:
            if len(params) != 6:
                message = (
                    f'If "{name}" is a sequence, it must have length 2, 3 or 6,'
                    f' not {len(params)}'
                )
                raise ValueError(message)
            for param_range in zip(params[::2], params[1::2]):
                self.parse_range(param_range, name, **kwargs)
        return tuple(params)

    @staticmethod
    def parse_range(
            nums_range: Union[TypeNumber, Tuple[TypeNumber, TypeNumber]],
            name: str,
            min_constraint: TypeNumber = None,
            max_constraint: TypeNumber = None,
            type_constraint: type = None,
            ) -> Tuple[TypeNumber, TypeNumber]:
        r"""Adapted from ``torchvision.transforms.RandomRotation``.

        Args:
            nums_range: Tuple of two numbers :math:`(n_{min}, n_{max})`,
                where :math:`n_{min} \leq n_{max}`.
                If a single positive number :math:`n` is provided,
                :math:`n_{min} = -n` and :math:`n_{max} = n`.
            name: Name of the parameter, so that an informative error message
                can be printed.
            min_constraint: Minimal value that :math:`n_{min}` can take,
                default is None, i.e. there is no minimal value.
            max_constraint: Maximal value that :math:`n_{max}` can take,
                default is None, i.e. there is no maximal value.
            type_constraint: Precise type that :math:`n_{max}` and
                :math:`n_{min}` must take.

        Returns:
            A tuple of two numbers :math:`(n_{min}, n_{max})`.

        Raises:
            ValueError: if :attr:`nums_range` is negative
            ValueError: if :math:`n_{max}` or :math:`n_{min}` is not a number
            ValueError: if :math:`n_{max} \lt n_{min}`
            ValueError: if :attr:`min_constraint` is not None and
                :math:`n_{min}` is smaller than :attr:`min_constraint`
            ValueError: if :attr:`max_constraint` is not None and
                :math:`n_{max}` is greater than :attr:`max_constraint`
            ValueError: if :attr:`type_constraint` is not None and
                :math:`n_{max}` and :math:`n_{max}` are not of type
                :attr:`type_constraint`.
        """
        if isinstance(nums_range, numbers.Number):  # single number given
            if nums_range < 0:
                raise ValueError(
                    f'If {name} is a single number,'
                    f' it must be positive, not {nums_range}')
            if min_constraint is not None and nums_range < min_constraint:
                raise ValueError(
                    f'If {name} is a single number, it must be greater'
                    f' than {min_constraint}, not {nums_range}'
                )
            if max_constraint is not None and nums_range > max_constraint:
                raise ValueError(
                    f'If {name} is a single number, it must be smaller'
                    f' than {max_constraint}, not {nums_range}'
                )
            if type_constraint is not None:
                if not isinstance(nums_range, type_constraint):
                    raise ValueError(
                        f'If {name} is a single number, it must be of'
                        f' type {type_constraint}, not {nums_range}'
                    )
            min_range = -nums_range if min_constraint is None else nums_range
            return (min_range, nums_range)

        try:
            min_value, max_value = nums_range
        except (TypeError, ValueError):
            raise ValueError(
                f'If {name} is not a single number, it must be'
                f' a sequence of len 2, not {nums_range}'
            )

        min_is_number = isinstance(min_value, numbers.Number)
        max_is_number = isinstance(max_value, numbers.Number)
        if not min_is_number or not max_is_number:
            message = (
                f'{name} values must be numbers, not {nums_range}')
            raise ValueError(message)

        if min_value > max_value:
            raise ValueError(
                f'If {name} is a sequence, the second value must be'
                f' equal or greater than the first, but it is {nums_range}')

        if min_constraint is not None and min_value < min_constraint:
            raise ValueError(
                f'If {name} is a sequence, the first value must be greater'
                f' than {min_constraint}, but it is {min_value}'
            )

        if max_constraint is not None and max_value > max_constraint:
            raise ValueError(
                f'If {name} is a sequence, the second value must be smaller'
                f' than {max_constraint}, but it is {max_value}'
            )

        if type_constraint is not None:
            min_type_ok = isinstance(min_value, type_constraint)
            max_type_ok = isinstance(max_value, type_constraint)
            if not min_type_ok or not max_type_ok:
                raise ValueError(
                    f'If "{name}" is a sequence, its values must be of'
                    f' type "{type_constraint}", not "{type(nums_range)}"'
                )
        return nums_range

    @staticmethod
    def parse_interpolation(interpolation: str) -> str:
        interpolation = interpolation.lower()
        is_string = isinstance(interpolation, str)
        supported_values = [key.name.lower() for key in Interpolation]
        is_supported = interpolation.lower() in supported_values
        if is_string and is_supported:
            return interpolation
        message = (
            f'Interpolation "{interpolation}" of type {type(interpolation)}'
            f' must be a string among the supported values: {supported_values}'
        )
        raise TypeError(message)

    @staticmethod
    def parse_probability(probability: float) -> float:
        is_number = isinstance(probability, numbers.Number)
        if not (is_number and 0 <= probability <= 1):
            message = (
                'Probability must be a number in [0, 1],'
                f' not {probability}'
            )
            raise ValueError(message)
        return probability

    @staticmethod
    def nib_to_sitk(data: TypeData, affine: TypeData) -> sitk.Image:
        return nib_to_sitk(data, affine)

    @staticmethod
    def sitk_to_nib(image: sitk.Image) -> Tuple[torch.Tensor, np.ndarray]:
        return sitk_to_nib(image)

    @property
    def name(self):
        return self.__class__.__name__

    def get_arguments(self):
        """
        Return a dictionary with the arguments that would be necessary to
        reproduce the transform exactly.
        """
        return {name: getattr(self, name) for name in self.args_names}

    def is_invertible(self):
        return hasattr(self, 'invert_transform')

    def inverse(self):
        if not self.is_invertible():
            raise RuntimeError(f'{self.name} is not invertible')
        new = copy.deepcopy(self)
        new.invert_transform = not self.invert_transform
        return new


class DataToSubject:
    def __init__(
            self,
            data: TypeTransformInput,
            keys: Optional[List[str]] = None,
            ):
        self.data = data
        self.keys = keys
        self.default_image_name = 'default_image_name'
        self.is_tensor = False
        self.is_array = False
        self.is_dict = False
        self.is_image = False
        self.is_sitk = False
        self.is_nib = False

    def get_subject(self):
        if isinstance(self.data, nib.Nifti1Image):
            tensor = self.data.get_fdata(dtype=np.float32)
            data = ScalarImage(tensor=tensor, affine=self.data.affine)
            subject = self._get_subject_from_image(data)
            self.is_nib = True
        elif isinstance(self.data, (np.ndarray, torch.Tensor)):
            subject = self._parse_tensor(self.data)
            self.is_array = isinstance(self.data, np.ndarray)
            self.is_tensor = True
        elif isinstance(self.data, Image):
            subject = self._get_subject_from_image(self.data)
            self.is_image = True
        elif isinstance(self.data, Subject):
            subject = self.data
        elif isinstance(self.data, sitk.Image):
            subject = self._get_subject_from_sitk_image(self.data)
            self.is_sitk = True
        elif isinstance(self.data, dict):  # e.g. Eisen or MONAI dicts
            if self.keys is None:
                message = (
                    'If input is a dictionary, a value for "keys" must be'
                    ' specified when instantiating the transform'
                )
                raise RuntimeError(message)
            subject = self._get_subject_from_dict(self.data, self.keys)
            self.is_dict = True
        else:
            raise ValueError(f'Input type not recognized: {type(self.data)}')
        self._parse_subject(subject)
        return subject

    def get_output(self, transformed):
        if self.is_tensor or self.is_sitk:
            image = transformed[self.default_image_name]
            transformed = image[DATA]
            if self.is_array:
                transformed = transformed.numpy()
            elif self.is_sitk:
                transformed = nib_to_sitk(image[DATA], image[AFFINE])
        elif self.is_image:
            transformed = transformed[self.default_image_name]
        elif self.is_dict:
            transformed = dict(transformed)
            for key, value in transformed.items():
                if isinstance(value, Image):
                    transformed[key] = value.data
        elif self.is_nib:
            image = transformed[self.default_image_name]
            data = image[DATA]
            if len(data) > 1:
                message = (
                    'Multichannel images not supported for input of type'
                    ' nibabel.nifti.Nifti1Image'
                )
                raise RuntimeError(message)
            transformed = nib.Nifti1Image(data[0].numpy(), image[AFFINE])
        return transformed

    @staticmethod
    def _parse_subject(subject: Subject) -> None:
        if not isinstance(subject, Subject):
            message = (
                'Input to a transform must be a tensor or an instance'
                f' of torchio.Subject, not "{type(subject)}"'
            )
            raise RuntimeError(message)

    def _parse_tensor(self, data: TypeData) -> Subject:
        if data.ndim != 4:
            message = (
                'The input must be a 4D tensor with dimensions'
                f' (channels, x, y, z) but it has shape {tuple(data.shape)}'
            )
            raise ValueError(message)
        return self._get_subject_from_tensor(data)

    def _get_subject_from_tensor(self, tensor: torch.Tensor) -> Subject:
        image = ScalarImage(tensor=tensor)
        return self._get_subject_from_image(image)

    def _get_subject_from_image(self, image: Image) -> Subject:
        subject = Subject({self.default_image_name: image})
        return subject

    @staticmethod
    def _get_subject_from_dict(
            data: dict,
            image_keys: List[str],
            ) -> Subject:
        subject_dict = {}
        for key, value in data.items():
            if key in image_keys:
                value = ScalarImage(tensor=value)
            subject_dict[key] = value
        return Subject(subject_dict)

    def _get_subject_from_sitk_image(self, image):
        tensor, affine = sitk_to_nib(image)
        image = ScalarImage(tensor=tensor, affine=affine)
        return self._get_subject_from_image(image)