|
|
import numpy as np |
|
|
import torch |
|
|
|
|
|
from typing import Dict |
|
|
from torch import Tensor |
|
|
|
|
|
from .utility import chance |
|
|
|
|
|
|
|
|
def remap_intesities( |
|
|
image: Tensor, |
|
|
bins: int = 256) -> Tensor: |
|
|
""" |
|
|
Remap image intensities to a random distribution. |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
image : torch.Tensor |
|
|
Image to be remapped |
|
|
bins : int, optional |
|
|
The number of intensity bins to use when remapping the image. |
|
|
The default value of 256 is appropriate for 8-bit images. |
|
|
|
|
|
Returns |
|
|
------- |
|
|
torch.Tensor |
|
|
Remapped image |
|
|
""" |
|
|
device = image.device |
|
|
|
|
|
image = image.type(torch.float32) - image.min() |
|
|
image /= image.max() |
|
|
image *= (bins - 1) |
|
|
image = image.type(torch.int64) |
|
|
|
|
|
|
|
|
|
|
|
samples = 2 |
|
|
radians = 1 |
|
|
|
|
|
|
|
|
noise = torch.ones(bins, device=device, dtype=torch.float32) |
|
|
for i in range(samples): |
|
|
low = np.random.uniform(-radians, 0) |
|
|
high = np.random.uniform(0, radians) |
|
|
noise *= torch.sin(torch.linspace(low, high, bins, |
|
|
device=device, dtype=torch.float32)) |
|
|
|
|
|
noise -= noise.min() |
|
|
noise /= noise.max() |
|
|
return noise[image] |
|
|
|
|
|
|
|
|
def densities_to_image(densities: Tensor) -> Tensor: |
|
|
""" |
|
|
Replace density values with random signal between 0 and 1. |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
densities : torch.Tensor |
|
|
Multi-channel image of class density (probability) values |
|
|
|
|
|
Returns |
|
|
------- |
|
|
torch.Tensor |
|
|
Synthetic image with the same geometry as the input `densities` tensor. |
|
|
""" |
|
|
dims = [1] * (densities.ndim - 1) |
|
|
intensities = torch.rand(densities.shape[0], *dims) |
|
|
return torch.sum(densities * intensities, axis=0).unsqueeze(0) |
|
|
|
|
|
|
|
|
def labels_to_image( |
|
|
labels: Tensor, |
|
|
intensity_ranges: Dict = None) -> Tensor: |
|
|
""" |
|
|
Replace segmentation labels with random signal. |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
labels : torch.Tensor |
|
|
Image of integer segmentation labels |
|
|
intensity_ranges : dict, optional |
|
|
A dictionary mapping label values to intensity range tuples (low, high). |
|
|
Intensity values will be randomly generated within the specified range |
|
|
for each label. If not provided, a random intensity will be chosen for |
|
|
each label. |
|
|
|
|
|
Returns |
|
|
------- |
|
|
torch.Tensor |
|
|
Synthetic image with the same shape as the input `labels` tensor. |
|
|
""" |
|
|
labels = labels.type(torch.int64) |
|
|
max_label = labels.max() + 1 |
|
|
if intensity_ranges is not None: |
|
|
mapping = torch.zeros(max_label, device=labels.device, dtype=torch.float32) |
|
|
for k, (low, high) in intensity_ranges.items(): |
|
|
mapping[k] = np.random.uniform(low, high) |
|
|
else: |
|
|
mapping = torch.rand(max_label, device=labels.device, dtype=torch.float32) |
|
|
image = mapping[labels] |
|
|
return image |
|
|
|
