File size: 8,871 Bytes
2260825
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import numpy as np
import PIL.Image

from .file_utils import _is_torch, is_torch_available


IMAGENET_DEFAULT_MEAN = [0.485, 0.456, 0.406]
IMAGENET_DEFAULT_STD = [0.229, 0.224, 0.225]


def is_torch_tensor(obj):
    return _is_torch(obj) if is_torch_available() else False


# In the future we can add a TF implementation here when we have TF models.
class ImageFeatureExtractionMixin:
    """
    Mixin that contain utilities for preparing image features.
    """

    def _ensure_format_supported(self, image):
        if not isinstance(image, (PIL.Image.Image, np.ndarray)) and not is_torch_tensor(image):
            raise ValueError(
                f"Got type {type(image)} which is not supported, only `PIL.Image.Image`, `np.array` and "
                "`torch.Tensor` are."
            )

    def to_pil_image(self, image, rescale=None):
        """
        Converts :obj:`image` to a PIL Image. Optionally rescales it and puts the channel dimension back as the last
        axis if needed.

        Args:
            image (:obj:`PIL.Image.Image` or :obj:`numpy.ndarray` or :obj:`torch.Tensor`):
                The image to convert to the PIL Image format.
            rescale (:obj:`bool`, `optional`):
                Whether or not to apply the scaling factor (to make pixel values integers between 0 and 255). Will
                default to :obj:`True` if the image type is a floating type, :obj:`False` otherwise.
        """
        self._ensure_format_supported(image)

        if is_torch_tensor(image):
            image = image.numpy()

        if isinstance(image, np.ndarray):
            if rescale is None:
                # rescale default to the array being of floating type.
                rescale = isinstance(image.flat[0], np.floating)
            # If the channel as been moved to first dim, we put it back at the end.
            if image.ndim == 3 and image.shape[0] in [1, 3]:
                image = image.transpose(1, 2, 0)
            if rescale:
                image = image * 255
            image = image.astype(np.uint8)
            return PIL.Image.fromarray(image)
        return image

    def to_numpy_array(self, image, rescale=None, channel_first=True):
        """
        Converts :obj:`image` to a numpy array. Optionally rescales it and puts the channel dimension as the first
        dimension.

        Args:
            image (:obj:`PIL.Image.Image` or :obj:`np.ndarray` or :obj:`torch.Tensor`):
                The image to convert to a NumPy array.
            rescale (:obj:`bool`, `optional`):
                Whether or not to apply the scaling factor (to make pixel values floats between 0. and 1.). Will
                default to :obj:`True` if the image is a PIL Image or an array/tensor of integers, :obj:`False`
                otherwise.
            channel_first (:obj:`bool`, `optional`, defaults to :obj:`True`):
                Whether or not to permute the dimensions of the image to put the channel dimension first.
        """
        self._ensure_format_supported(image)

        if isinstance(image, PIL.Image.Image):
            image = np.array(image)

        if is_torch_tensor(image):
            image = image.numpy()

        if rescale is None:
            rescale = isinstance(image.flat[0], np.integer)

        if rescale:
            image = image.astype(np.float32) / 255.0

        if channel_first:
            image = image.transpose(2, 0, 1)

        return image

    def normalize(self, image, mean, std):
        """
        Normalizes :obj:`image` with :obj:`mean` and :obj:`std`. Note that this will trigger a conversion of
        :obj:`image` to a NumPy array if it's a PIL Image.

        Args:
            image (:obj:`PIL.Image.Image` or :obj:`np.ndarray` or :obj:`torch.Tensor`):
                The image to normalize.
            mean (:obj:`List[float]` or :obj:`np.ndarray` or :obj:`torch.Tensor`):
                The mean (per channel) to use for normalization.
            std (:obj:`List[float]` or :obj:`np.ndarray` or :obj:`torch.Tensor`):
                The standard deviation (per channel) to use for normalization.
        """
        self._ensure_format_supported(image)

        if isinstance(image, PIL.Image.Image):
            image = self.to_numpy_array(image)

        if isinstance(image, np.ndarray):
            if not isinstance(mean, np.ndarray):
                mean = np.array(mean).astype(image.dtype)
            if not isinstance(std, np.ndarray):
                std = np.array(std).astype(image.dtype)
        elif is_torch_tensor(image):
            import torch

            if not isinstance(mean, torch.Tensor):
                mean = torch.tensor(mean)
            if not isinstance(std, torch.Tensor):
                std = torch.tensor(std)

        if image.ndim == 3 and image.shape[0] in [1, 3]:
            return (image - mean[:, None, None]) / std[:, None, None]
        else:
            return (image - mean) / std

    def resize(self, image, size, resample=PIL.Image.BILINEAR):
        """
        Resizes :obj:`image`. Note that this will trigger a conversion of :obj:`image` to a PIL Image.

        Args:
            image (:obj:`PIL.Image.Image` or :obj:`np.ndarray` or :obj:`torch.Tensor`):
                The image to resize.
            size (:obj:`int` or :obj:`Tuple[int, int]`):
                The size to use for resizing the image.
            resample (:obj:`int`, `optional`, defaults to :obj:`PIL.Image.BILINEAR`):
                The filter to user for resampling.
        """
        self._ensure_format_supported(image)

        if not isinstance(size, tuple):
            size = (size, size)
        if not isinstance(image, PIL.Image.Image):
            image = self.to_pil_image(image)

        return image.resize(size, resample=resample)

    def center_crop(self, image, size):
        """
        Crops :obj:`image` to the given size using a center crop. Note that if the image is too small to be cropped to
        the size given, it will be padded (so the returned result has the size asked).

        Args:
            image (:obj:`PIL.Image.Image` or :obj:`np.ndarray` or :obj:`torch.Tensor`):
                The image to resize.
            size (:obj:`int` or :obj:`Tuple[int, int]`):
                The size to which crop the image.
        """
        self._ensure_format_supported(image)
        if not isinstance(size, tuple):
            size = (size, size)

        # PIL Image.size is (width, height) but NumPy array and torch Tensors have (height, width)
        image_shape = (image.size[1], image.size[0]) if isinstance(image, PIL.Image.Image) else image.shape[-2:]
        top = (image_shape[0] - size[0]) // 2
        bottom = top + size[0]  # In case size is odd, (image_shape[0] + size[0]) // 2 won't give the proper result.
        left = (image_shape[1] - size[1]) // 2
        right = left + size[1]  # In case size is odd, (image_shape[1] + size[1]) // 2 won't give the proper result.

        # For PIL Images we have a method to crop directly.
        if isinstance(image, PIL.Image.Image):
            return image.crop((left, top, right, bottom))

        # Check if all the dimensions are inside the image.
        if top >= 0 and bottom <= image_shape[0] and left >= 0 and right <= image_shape[1]:
            return image[..., top:bottom, left:right]

        # Otherwise, we may need to pad if the image is too small. Oh joy...
        new_shape = image.shape[:-2] + (max(size[0], image_shape[0]), max(size[1], image_shape[1]))
        if isinstance(image, np.ndarray):
            new_image = np.zeros_like(image, shape=new_shape)
        elif is_torch_tensor(image):
            new_image = image.new_zeros(new_shape)

        top_pad = (new_shape[-2] - image_shape[0]) // 2
        bottom_pad = top_pad + image_shape[0]
        left_pad = (new_shape[-1] - image_shape[1]) // 2
        right_pad = left_pad + image_shape[1]
        new_image[..., top_pad:bottom_pad, left_pad:right_pad] = image

        top += top_pad
        bottom += top_pad
        left += left_pad
        right += left_pad

        return new_image[
            ..., max(0, top) : min(new_image.shape[-2], bottom), max(0, left) : min(new_image.shape[-1], right)
        ]