File size: 6,682 Bytes
c4c7cee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
216
217
218
219
220
221
222
223
224
"""
Adapted from https://github.com/nv-nguyen/template-pose/blob/main/src/utils/augmentation.py
"""

from torchvision import transforms
from PIL import ImageEnhance, ImageFilter, Image
import numpy as np
import random
import logging
from torchvision.transforms import RandomResizedCrop, ToTensor


class PillowRGBAugmentation:
    def __init__(self, pillow_fn, p, factor_interval):
        self._pillow_fn = pillow_fn
        self.p = p
        self.factor_interval = factor_interval

    def __call__(self, PIL_image):
        if random.random() <= self.p:
            factor = random.uniform(*self.factor_interval)
            if PIL_image.mode != "RGB":
                logging.warning(
                    f"Error when apply data aug, image mode: {PIL_image.mode}"
                )
                imgs = imgs.convert("RGB")
                logging.warning(f"Success to change to {PIL_image.mode}")
            PIL_image = (self._pillow_fn(PIL_image).enhance(factor=factor)).convert(
                "RGB"
            )
        return PIL_image


class PillowSharpness(PillowRGBAugmentation):
    def __init__(
        self,
        p=0.3,
        factor_interval=(0, 40.0),
    ):
        super().__init__(
            pillow_fn=ImageEnhance.Sharpness,
            p=p,
            factor_interval=factor_interval,
        )


class PillowContrast(PillowRGBAugmentation):
    def __init__(
        self,
        p=0.3,
        factor_interval=(0.5, 1.6),
    ):
        super().__init__(
            pillow_fn=ImageEnhance.Contrast,
            p=p,
            factor_interval=factor_interval,
        )


class PillowBrightness(PillowRGBAugmentation):
    def __init__(
        self,
        p=0.5,
        factor_interval=(0.5, 2.0),
    ):
        super().__init__(
            pillow_fn=ImageEnhance.Brightness,
            p=p,
            factor_interval=factor_interval,
        )


class PillowColor(PillowRGBAugmentation):
    def __init__(
        self,
        p=1,
        factor_interval=(0.0, 20.0),
    ):
        super().__init__(
            pillow_fn=ImageEnhance.Color,
            p=p,
            factor_interval=factor_interval,
        )


class PillowBlur:
    def __init__(self, p=0.4, factor_interval=(1, 3)):
        self.p = p
        self.k = random.randint(*factor_interval)

    def __call__(self, PIL_image):
        if random.random() <= self.p:
            PIL_image = PIL_image.filter(ImageFilter.GaussianBlur(self.k))
        return PIL_image


class NumpyGaussianNoise:
    def __init__(self, p, factor_interval=(0.01, 0.3)):
        self.noise_ratio = random.uniform(*factor_interval)
        self.p = p

    def __call__(self, img):
        if random.random() <= self.p:
            img = np.copy(img)
            noisesigma = random.uniform(0, self.noise_ratio)
            gauss = np.random.normal(0, noisesigma, img.shape) * 255
            img = img + gauss

            img[img > 255] = 255
            img[img < 0] = 0
        return Image.fromarray(np.uint8(img))


class StandardAugmentation:
    def __init__(
        self, names, brightness, contrast, sharpness, color, blur, gaussian_noise
    ):
        self.brightness = brightness
        self.contrast = contrast
        self.sharpness = sharpness
        self.color = color
        self.blur = blur
        self.gaussian_noise = gaussian_noise

        # define a dictionary of augmentation functions to be applied
        self.names = names.split(",")
        self.augmentations = {
            "brightness": self.brightness,
            "contrast": self.contrast,
            "sharpness": self.sharpness,
            "color": self.color,
            "blur": self.blur,
            "gaussian_noise": self.gaussian_noise,
        }

    def __call__(self, img):
        for name in self.names:
            img = self.augmentations[name](img)
        return img


class GeometricAugmentation:
    def __init__(
        self,
        names,
        random_resized_crop,
        random_horizontal_flip,
        random_vertical_flip,
        random_rotation,
    ):
        self.random_resized_crop = random_resized_crop
        self.random_horizontal_flip = random_horizontal_flip
        self.random_vertical_flip = random_vertical_flip
        self.random_rotation = random_rotation
        self.names = names.split(",")

        self.augmentations = {
            "random_resized_crop": self.random_resized_crop,
            "random_horizontal_flip": self.random_horizontal_flip,
            "random_vertical_flip": self.random_vertical_flip,
            "random_rotation": self.random_rotation,
        }

    def __call__(self, img):
        for name in self.names:
            img = self.augmentations[name](img)
        return img


class ImageAugmentation:
    def __init__(
        self, names, clip_transform, standard_augmentation, geometric_augmentation
    ):
        self.clip_transform = clip_transform
        self.standard_augmentation = standard_augmentation
        self.geometric_augmentation = geometric_augmentation
        self.names = names.split(",")
        self.transforms = {
            "clip_transform": self.clip_transform,
            "standard_augmentation": self.standard_augmentation,
            "geometric_augmentation": self.geometric_augmentation,
        }
        print(f"Image augmentation: {self.names}")

    def __call__(self, img):
        for name in self.names:
            img = self.transforms[name](img)
        return img


if __name__ == "__main__":
    # sanity check
    import glob
    import torchvision.transforms as transforms
    from torchvision.utils import save_image
    from omegaconf import DictConfig, OmegaConf
    from hydra.utils import instantiate
    import torch
    from PIL import Image

    augmentation_config = OmegaConf.load(
        "./configs/dataset/train_transform/augmentation.yaml"
    )
    augmentation_config.names = "standard_augmentation,geometric_augmentation"
    augmentation_transform = instantiate(augmentation_config)
    img_paths = glob.glob("./datasets/osv5m/test/images/*.jpg")

    num_try = 20
    num_try_per_image = 8
    num_imgs = 8

    for idx in range(num_try):
        imgs = []
        for idx_img in range(num_imgs):
            img = Image.open(img_paths[idx_img])
            for idx_try in range(num_try_per_image):
                if idx_try == 0:
                    imgs.append(ToTensor()(img.resize((224, 224))))
                img_aug = augmentation_transform(img.copy())
                img_aug = ToTensor()(img_aug)
                imgs.append(img_aug)
        imgs = torch.stack(imgs)
        save_image(imgs, f"augmentation_{idx:03d}.png", nrow=9)