|
1
|
|
|
import numpy as np |
|
2
|
|
|
from numpy.typing import NDArray |
|
3
|
|
|
from typing import Tuple |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
def reshape_image(image: NDArray, shape: Tuple[int, ...]) -> NDArray: |
|
7
|
|
|
return np.reshape(image, shape) |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
def subtract(image: NDArray, array: NDArray) -> NDArray: |
|
11
|
|
|
"""Normalize the input image. |
|
12
|
|
|
|
|
13
|
|
|
Args: |
|
14
|
|
|
image (NDArray): [description] |
|
15
|
|
|
|
|
16
|
|
|
Raises: |
|
17
|
|
|
ShapeMissmatchError: in case of ValueError due to numpy broadcasting failing |
|
18
|
|
|
|
|
19
|
|
|
Returns: |
|
20
|
|
|
NDArray: [description] |
|
21
|
|
|
""" |
|
22
|
|
|
try: |
|
23
|
|
|
return image - array |
|
24
|
|
|
except ValueError as numpy_broadcast_error: |
|
25
|
|
|
raise ShapeMissmatchError( |
|
26
|
|
|
f'Expected arrays with matching shapes.') from numpy_broadcast_error |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
class ShapeMissmatchError(Exception): pass |
|
30
|
|
|
|
|
31
|
|
|
|
|
32
|
|
|
def noisy(image: NDArray, ratio: float) -> NDArray: |
|
33
|
|
|
"""Generates a noisy image by adding random noise to the content_image""" |
|
34
|
|
|
if ratio < 0 or 1 < ratio: |
|
35
|
|
|
raise InvalidRatioError('Expected a ratio value x such that 0 <= x <= 1') |
|
36
|
|
|
|
|
37
|
|
|
prod_shape = image.shape |
|
38
|
|
|
# assert prod_shape == (1, self.config.image_height, self.config.image_width, self.config.color_channels) |
|
39
|
|
|
|
|
40
|
|
|
noise_image = np.random.uniform(-20, 20, prod_shape).astype('float32') |
|
41
|
|
|
|
|
42
|
|
|
# Set the input_image to be a weighted average of the content_image and a noise_image |
|
43
|
|
|
return noise_image * ratio + image * (1 - ratio) |
|
44
|
|
|
|
|
45
|
|
|
|
|
46
|
|
|
class InvalidRatioError(Exception): pass |
|
47
|
|
|
|
|
48
|
|
|
|
|
49
|
|
|
class ImageDTypeConverter: |
|
50
|
|
|
|
|
51
|
|
|
bit_2_data_type = {8: np.uint8} |
|
52
|
|
|
|
|
53
|
|
|
def __call__(self, image: NDArray): |
|
54
|
|
|
return self._convert_to_uint8(image) |
|
55
|
|
|
|
|
56
|
|
|
def _convert_to_uint8(self, im): |
|
57
|
|
|
bitdepth = 8 |
|
58
|
|
|
out_type = type(self).bit_2_data_type[bitdepth] |
|
59
|
|
|
mi = np.nanmin(im) |
|
60
|
|
|
ma = np.nanmax(im) |
|
61
|
|
|
if not np.isfinite(mi): |
|
62
|
|
|
raise ValueError("Minimum image value is not finite") |
|
63
|
|
|
if not np.isfinite(ma): |
|
64
|
|
|
raise ValueError("Maximum image value is not finite") |
|
65
|
|
|
if ma == mi: |
|
66
|
|
|
return im.astype(out_type) |
|
67
|
|
|
|
|
68
|
|
|
# Make float copy before we scale |
|
69
|
|
|
im = im.astype("float64") |
|
70
|
|
|
# Scale the values between 0 and 1 then multiply by the max value |
|
71
|
|
|
im = (im - mi) / (ma - mi) * (np.power(2.0, bitdepth) - 1) + 0.499999999 |
|
72
|
|
|
assert np.nanmin(im) >= 0 |
|
73
|
|
|
assert np.nanmax(im) < np.power(2.0, bitdepth) |
|
74
|
|
|
im = im.astype(out_type) |
|
75
|
|
|
return im |
|
76
|
|
|
|
|
77
|
|
|
|
|
78
|
|
|
convert_to_uint8 = ImageDTypeConverter() |
|
79
|
|
|
|
boromir674 /
neural-style-transfer
