Datasets:

ArXiv:
PRISM / src /voxynth /transform.py
emad2001's picture
Upload folder using huggingface_hub
36fdbcf verified
import numpy as np
import torch
from typing import List
from torch import Tensor
from .utility import chance
from .utility import grid_coordinates
from .noise import perlin
def resize(
image : Tensor,
scale_factor : float or List[float] = None,
shape : List[int] = None,
nearest : bool = False) -> Tensor:
"""
Resize an image with the option of scaling and/or setting to a new shape.
Parameters:
-----------
image: torch.Tensor
An input tensor with shape (C, H, W[, D]) to resize.
scale_factor: float or List[float], optional
Multiplicative factor(s) for scaling the input tensor. If a float, then the same
scale factor is applied to all spatial dimensions. If a tuple, then the scaling
factor for each dimension should be provided.
shape: List[int], optional
Target shape of the output tensor.
nearest: bool, optional
If True, use nearest neighbor interpolation. Otherwise, use linear interpolation.
Returns:
--------
torch.Tensor:
The resized tensor with the shape specified by `shape` or scaled by `scale_factor`.
"""
ndim = image.ndim - 1
# scale the image if the scale factor is provided
if scale_factor is not None and scale_factor != 1:
# compute target shape based on the scale factor
target_shape = [int(s * scale_factor + 0.5) for s in image.shape[1:]]
# convert image to float32 if it's not already to enable interpolation
# if using nearest interpolation, save the original dtype to convert back later
reset_type = None
if not torch.is_floating_point(image):
if nearest:
reset_type = image.dtype
image = image.type(torch.float32)
# determine interpolation mode based on ndim and interpolation type
linear = 'trilinear' if image.ndim - 1 == 3 else 'bilinear'
mode = 'nearest' if nearest else linear
# apply interpolation to the image
if nearest:
image = torch.nn.functional.interpolate(image.unsqueeze(0), target_shape, mode=mode)
else:
image = torch.nn.functional.interpolate(image.unsqueeze(0), target_shape, mode=mode)
image = image.squeeze(0)
# convert image back to its original dtype if necessary
if reset_type is not None:
image = image.type(reset_type)
if shape is not None:
# compute padding for each spatial dimension
padding = []
baseshape = image.shape[1:]
for d in range(ndim):
diff = shape[d] - baseshape[d]
if diff > 0:
half = diff / 2
a, b = int(np.floor(half)), int(np.ceil(half))
padding.extend([a, b])
else:
padding.extend([0, 0])
# apply padding to the image
padding.reverse()
image = torch.nn.functional.pad(image, padding)
# compute slice to remove excess dimensions
slicing = [slice(0, image.shape[0])]
baseshape = image.shape[1:]
for d in range(ndim):
diff = baseshape[d] - shape[d]
if diff > 0:
half = diff / 2
a, b = int(np.floor(half)), int(np.ceil(half))
slicing.append(slice(a, baseshape[d] - b))
else:
slicing.append(slice(0, baseshape[d]))
# apply slice to remove excess dimensions
image = image[tuple(slicing)]
return image
def random_flip(dim : int, *args, prob : float = 0.5):
"""
Randomly flips an image (or set of images) along the given dimension.
Parameters
----------
dim : int
The dimension along which to flip. Note that the first dimension
is the channel dimension.
*args : torch.Tensor
The image(s) to flip.
prob : float
The probability of flipping the image(s).
Returns
-------
torch.Tensor or tuple[torch.Tensor]
The flipped image(s).
"""
result = tuple([arg.flip([dim]) for arg in args]) if chance(prob) else args
if len(args) == 1:
return result[0]
return result
def compose_affine(
ndim : int,
translation : Tensor = None,
rotation : Tensor = None,
scale : Tensor = None,
shear : Tensor = None,
degrees : bool = True,
device : torch.device = None) -> Tensor:
"""
Composes an affine matrix from a set of translation, rotation, scale,
and shear transform components.
Parameters
----------
ndim (int):
The number of dimensions of the affine matrix. Must be 2 or 3.
translation : Tensor, optional
The translation vector. Must be a vector of size `ndim`.
rotation : Tensor, optional
The rotation angles. Must be a scalar value for 2D affine matrices,
and a tensor of size 3 for 3D affine matrices.
scale : Tensor, optional
The scaling factor. Can be scalar or vector of size `ndim`.
shear : Tensor, optional
The shearing factor. Must be a scalar value for 2D affine matrices,
and a tensor of size 3 for 3D affine matrices.
degrees : bool, optional
Whether to interpret the rotation angles as degrees.
device : torch.device, optional
The device of the returned matrix.
Returns
-------
Tensor
The composed affine matrix, as a tensor of shape `(ndim + 1, ndim + 1)`.
"""
if ndim not in (2, 3):
raise ValueError(f'affine transform must be 2D or 3D, got ndim {ndim}')
# check translation
translation = torch.zeros(ndim) if translation is None else torch.as_tensor(translation)
if len(translation) != ndim:
raise ValueError(f'translation must be of shape ({ndim},)')
# check rotation angles
expected = 3 if ndim == 3 else 1
rotation = torch.zeros(expected) if rotation is None else torch.as_tensor(rotation)
if rotation.ndim == 0 and ndim == 3 or rotation.ndim != 0 and rotation.shape[0] != expected:
raise ValueError(f'rotation must be of shape ({expected},)')
# check scaling factor
scale = torch.ones(ndim) if scale is None else torch.as_tensor(scale)
if scale.ndim == 0:
scale = scale.repeat(ndim)
if scale.shape[0] != ndim:
raise ValueError(f'scale must be of size {ndim}')
# check shearing
expected = 3 if ndim == 3 else 1
shear = torch.zeros(expected) if shear is None else torch.as_tensor(shear)
if shear.ndim == 0:
shear = shear.view(1)
if shear.shape[0] != expected:
raise ValueError(f'shear must be of shape ({expected},)')
# start from translation
T = torch.eye(ndim + 1, dtype=torch.float64)
T[:ndim, -1] = translation
# rotation matrix
R = torch.eye(ndim + 1, dtype=torch.float64)
R[:ndim, :ndim] = angles_to_rotation_matrix(rotation, degrees=degrees)
# scaling
Z = torch.diag(torch.cat([scale, torch.ones(1, dtype=torch.float64)]))
# shear matrix
S = torch.eye(ndim + 1, dtype=torch.float64)
S[0][1] = shear[0]
if ndim == 3:
S[0][2] = shear[1]
S[1][2] = shear[2]
# compose component matrices
matrix = T @ R @ Z @ S
return torch.as_tensor(matrix, dtype=torch.float32, device=device)
def angles_to_rotation_matrix(
rotation : Tensor,
degrees : bool = True) -> Tensor:
"""
Compute a rotation matrix from the given rotation angles.
Parameters
----------
rotation : Tensor
A tensor containing the rotation angles. If `degrees` is True, the angles
are in degrees, otherwise they are in radians.
degrees : bool, optional
Whether to interpret the rotation angles as degrees.
Returns
-------
Tensor
The computed `(ndim + 1, ndim + 1)` rotation matrix.
"""
if degrees:
rotation = torch.deg2rad(rotation)
# scalar value allowed for 2D transforms
rotation = torch.as_tensor(rotation)
if rotation.ndim == 0:
rotation = rotation.view(1)
num_angles = len(rotation)
# build the matrix
if num_angles == 1:
c, s = torch.cos(rotation[0]), torch.sin(rotation[0])
matrix = torch.tensor([[c, -s], [s, c]], dtype=torch.float64)
elif num_angles == 3:
c, s = torch.cos(rotation[0]), torch.sin(rotation[0])
rx = torch.tensor([[1, 0, 0], [0, c, s], [0, -s, c]], dtype=torch.float64)
c, s = torch.cos(rotation[1]), torch.sin(rotation[1])
ry = torch.tensor([[c, 0, s], [0, 1, 0], [-s, 0, c]], dtype=torch.float64)
c, s = torch.cos(rotation[2]), torch.sin(rotation[2])
rz = torch.tensor([[c, s, 0], [-s, c, 0], [0, 0, 1]], dtype=torch.float64)
matrix = rx @ ry @ rz
else:
raise ValueError(f'expected 1 (2D) or 3 (3D) rotation angles, got {num_angles}')
return matrix.to(rotation.device)
def random_affine(
ndim: int,
max_translation: float = 0,
max_rotation: float = 0,
max_scaling: float = 1,
device: torch.device = None) -> Tensor:
"""
TODOC
"""
#
translation_range = sorted([-max_translation, max_translation])
translation = np.random.uniform(*translation_range, size=ndim)
#
rotation_range = sorted([-max_rotation, max_rotation])
rotation = np.random.uniform(*rotation_range, size=(1 if ndim == 2 else 3))
#
if max_scaling < 1:
raise ValueError('max scaling to random affine cannot be less than 1, '
'see function doc for more info')
inv = np.random.choice([-1, 1], size=ndim)
scale = np.random.uniform(1, max_scaling, size=ndim) ** inv
# compose from random paramters
aff = compose_affine(
ndim=ndim,
translation=translation,
rotation=rotation,
scale=scale,
device=device)
return aff
def affine_to_displacement_field(
affine : Tensor,
meshgrid : Tensor,
rotate_around_center : bool = True) -> Tensor:
"""
Convert an affine transformation matrix to a dense displacement field.
Parameters
----------
affine : Tensor
Affine transformation matrix.
meshgrid : Tensor
The meshgrid tensor of shape `(W, H[, D], N)`, where N is the spatial dimensionality.
rotate_around_center : bool, optional
If True, the rotation will be around the center of the image, otherwise around the origin.
Returns
-------
Tensor
The generated displacement field of shape `meshgrid.shape[:-1]`.
"""
ndim = meshgrid.shape[-1]
shape = meshgrid.shape[:-1]
# if rotate_around_center is enabled, adjust the meshgrid so that the rotation
# is around the center of the image instead of the origin
grid = meshgrid.clone() if rotate_around_center else meshgrid
if rotate_around_center:
for d in range(ndim):
grid[..., d] -= (shape[d] - 1) / 2
# convert the meshgrid to homogeneous coordinates by appending a column of ones
coords = grid.view(-1, ndim)
ones = torch.ones((coords.shape[-2], 1), device=meshgrid.device)
coords = torch.cat([coords, ones], dim=-1)
# apply the affine transformation to the coordinates to get the shift vector
shift = (affine @ coords.T)[:ndim].T
# reshape the shift vector to match the shape of the meshgrid and subtract
# the original meshgrid to get the displacement field
shift = shift.view(*shape, ndim) - grid
return shift
def integrate_displacement_field(
disp : Tensor,
steps : int,
meshgrid : Tensor = None) -> Tensor:
"""
TODOC
"""
if meshgrid is None:
meshgrid = grid_coordinates(disp.shape[:-1], device=disp.device)
if steps == 0:
return disp
disp = disp / (2 ** steps)
for _ in range(steps):
disp += spatial_transform(disp.movedim(-1, 0), disp, meshgrid=meshgrid).movedim(0, -1)
return disp
def random_displacement_field(
shape : List[int],
smoothing : float = 10,
magnitude : float = 10,
integrations : int = 0,
voxsize : float = 1,
meshgrid : Tensor = None,
device: torch.device = None) -> Tensor:
"""
TODOC
"""
smoothing = smoothing / voxsize
magnitude = magnitude / voxsize
ndim = len(shape)
disp = [perlin(shape, smoothing, magnitude, device=device) for i in range(ndim)]
disp = torch.stack(disp, dim=-1)
if integrations > 0:
disp = integrate_displacement_field(disp, integrations, meshgrid)
return disp
def displacement_field_to_coords(disp, meshgrid=None) -> Tensor:
"""
TODOC
"""
if meshgrid is None:
meshgrid = grid_coordinates(disp.shape[:-1], device=disp.device)
shape = disp.shape[:-1]
ndim = disp.shape[-1]
coords = (meshgrid + disp)
for d in range(ndim):
coords[..., d] *= 2 / (shape[d] - 1)
coords[..., d] -= 1
coords = coords.flip(-1)
return coords
def coords_to_displacement_field(coords, meshgrid=None) -> Tensor:
"""
TODOC
"""
if meshgrid is None:
meshgrid = grid_coordinates(coords.shape[:-1], device=coords.device)
raise NotImplementedError('coords_to_displacement_field is not yet implemented. '
'contact andrew if you get this... or implement it :)')
def spatial_transform(
image : Tensor,
trf : Tensor,
method : str = 'linear',
isdisp : bool = True,
meshgrid : Tensor = None,
rotate_around_center : bool = True) -> Tensor:
"""
TODOC
"""
if trf is None:
return image
if trf.ndim == 2:
if meshgrid is None:
meshgrid = grid_coordinates(image.shape[1:], device=image.device)
trf = torch.linalg.inv(trf)
trf = affine_to_displacement_field(trf, meshgrid,
rotate_around_center=rotate_around_center)
isdisp = True
#
if isdisp:
trf = displacement_field_to_coords(trf, meshgrid=meshgrid)
#
method = 'bilinear' if method == 'linear' else method
#
reset_type = None
if not torch.is_floating_point(image):
if method == 'nearest':
reset_type = image.dtype
image = image.type(torch.float32)
#
image = image.unsqueeze(0)
trf = trf.unsqueeze(0)
interped = torch.nn.functional.grid_sample(image, trf, align_corners=True, mode=method)
interped = interped.squeeze(0)
#
if reset_type is not None:
interped = interped.type(reset_type)
return interped
def random_transform(
shape : List[int],
affine_probability : float = 1.0,
max_translation : float = 5.0,
max_rotation : float = 5.0,
max_scaling : float = 1.1,
warp_probability : float = 1.0,
warp_integrations : int = 5,
warp_smoothing_range : List[int] = [10, 20],
warp_magnitude_range : List[int] = [1, 2],
voxsize : int = 1,
device : torch.device = None,
isdisp : bool = True,
) -> Tensor:
"""
TODOC
"""
ndim = len(shape)
trf = None
# generate a random affine
if chance(affine_probability):
# compute meshgrid
meshgrid = grid_coordinates(shape, device=device)
# scale max translation value so that it correctly corresponds to mm
max_translation = max_translation / voxsize
matrix = random_affine(
ndim=ndim,
max_translation=max_translation,
max_rotation=max_rotation,
max_scaling=max_scaling,
device=device)
trf = affine_to_displacement_field(matrix, meshgrid)
# generate a nonlinear transform
if chance(warp_probability):
disp = random_displacement_field(
shape=shape,
smoothing=np.random.uniform(*warp_smoothing_range),
magnitude=np.random.uniform(*warp_magnitude_range),
integrations=warp_integrations,
voxsize=voxsize,
device=device)
# merge with the affine transform if necessary
if trf is None:
trf = disp
else:
trf += spatial_transform(disp.movedim(-1, 0), trf, meshgrid=meshgrid).movedim(0, -1)
# convert to coordinates if specified
if trf is not None and not isdisp:
trf = displacement_field_to_coords(trf)
return trf