Skip to content

Augmentation layers and models

Implementations of layers for augmenting images

Augmentation

Bases: Layer

Base augmentation class

Used to create Random Augmentation layers

Methods:

random_execute: returns boolean variable based on given probability. Determines whether the layer should activate on given batch augment_single_image: abstract method. Should be implemented by child classes. Distorts a single image random_augment_single_image: randomly augment an image with given probability call: run the layer on a batch of images

Source code in conftrainer/augmentation/layers.py 
 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
class Augmentation(Layer):
    """
    Base augmentation class

    Used to create Random Augmentation layers

    Methods:
    --------
    random_execute:
        returns boolean variable based on given probability. Determines whether the layer should
     activate on given batch
    augment_single_image:
        abstract method. Should be implemented by child classes. Distorts a single image
    random_augment_single_image:
        randomly augment an image with given probability
    call:
        run the layer on a batch of images
    """

    def __init__(self, factor, shape=None, **kwargs):
        super().__init__(**kwargs)
        self.shape = shape
        self.factor = factor

    @staticmethod
    def random_execute(probability):
        """
        Return boolean variable based on given probability

        Generate random float from 0 to 1 and return whether it's smaller than given float

        Parameters
        ----------
        probability: float
            probability to return True

        Returns
        -------
        out: bool
            whether generated value is smaller than given prob

        Raises
        ------
        NotImplementedError: if the child class doesn't implement this method
        """

        return tf.random.uniform([], minval=0, maxval=1) < probability

    def augment_single_image(self, image):
        """
        Abstract method for augmenting a single image.

        Should be implemented by child classes

        Parameters
        ----------
        image: tf.Tensor
            image to distort

        Returns
        -------
        out: tf.Tensor
            distorted image
        """

        raise NotImplementedError

    def random_augment_single_image(self, image):
        """
        Randomly augment a single image with given probabilty

        Parameters
        ----------
        image: tf.Tensor
            image to augment

        Returns
        --------
        out: tf.Tensor

        """

        if self.random_execute(self.factor):
            augmented = self.augment_single_image(image=image)
            return tf.clip_by_value(augmented, clip_value_min=0, clip_value_max=255)
        return image

    def call(self, inputs, training):
        """
        Randomly run the layer on given inputs

        Parameters
        ----------
        inputs: tf.Tensor
                batch of input images
        training: bool
                whether to run the layer in training or inference mode. The layer is inactive in
                inference mode and returns the inputs as is

        Returns
        -------
        out: tf.Tensor
            distorted version of images with given probability, original ones otherwise
        """
        if training:
            inputs = tf.image.convert_image_dtype(inputs, dtype=self.compute_dtype)
            return tf.map_fn(fn=self.random_augment_single_image, elems=inputs,
                             fn_output_signature=self.compute_dtype)
        return inputs

random_execute(probability) staticmethod

Return boolean variable based on given probability

Generate random float from 0 to 1 and return whether it's smaller than given float

Parameters:

Name Type Description Default
probability

probability to return True

 required

Returns:

Name Type Description
out bool

whether generated value is smaller than given prob

Raises:

Type Description
NotImplementedError: if the child class doesn't implement this method
Source code in conftrainer/augmentation/layers.py 
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
@staticmethod
def random_execute(probability):
    """
    Return boolean variable based on given probability

    Generate random float from 0 to 1 and return whether it's smaller than given float

    Parameters
    ----------
    probability: float
        probability to return True

    Returns
    -------
    out: bool
        whether generated value is smaller than given prob

    Raises
    ------
    NotImplementedError: if the child class doesn't implement this method
    """

    return tf.random.uniform([], minval=0, maxval=1) < probability

augment_single_image(image)

Abstract method for augmenting a single image.

Should be implemented by child classes

Parameters:

Name Type Description Default
image

image to distort

 required

Returns:

Name Type Description
out tf.Tensor

distorted image
Source code in conftrainer/augmentation/layers.py 
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
def augment_single_image(self, image):
    """
    Abstract method for augmenting a single image.

    Should be implemented by child classes

    Parameters
    ----------
    image: tf.Tensor
        image to distort

    Returns
    -------
    out: tf.Tensor
        distorted image
    """

    raise NotImplementedError

random_augment_single_image(image)

Randomly augment a single image with given probabilty

Parameters:

Name Type Description Default
image

image to augment

 required

Returns:

Name Type Description
out tf.Tensor
Source code in conftrainer/augmentation/layers.py 
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
def random_augment_single_image(self, image):
    """
    Randomly augment a single image with given probabilty

    Parameters
    ----------
    image: tf.Tensor
        image to augment

    Returns
    --------
    out: tf.Tensor

    """

    if self.random_execute(self.factor):
        augmented = self.augment_single_image(image=image)
        return tf.clip_by_value(augmented, clip_value_min=0, clip_value_max=255)
    return image

call(inputs, training)

Randomly run the layer on given inputs

Parameters:

Name Type Description Default
inputs

batch of input images

 required
training

whether to run the layer in training or inference mode. The layer is inactive in inference mode and returns the inputs as is

 required

Returns:

Name Type Description
out tf.Tensor

distorted version of images with given probability, original ones otherwise
Source code in conftrainer/augmentation/layers.py 
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
def call(self, inputs, training):
    """
    Randomly run the layer on given inputs

    Parameters
    ----------
    inputs: tf.Tensor
            batch of input images
    training: bool
            whether to run the layer in training or inference mode. The layer is inactive in
            inference mode and returns the inputs as is

    Returns
    -------
    out: tf.Tensor
        distorted version of images with given probability, original ones otherwise
    """
    if training:
        inputs = tf.image.convert_image_dtype(inputs, dtype=self.compute_dtype)
        return tf.map_fn(fn=self.random_augment_single_image, elems=inputs,
                         fn_output_signature=self.compute_dtype)
    return inputs

RandomColorJitter

Bases: Augmentation

Randomly Jitter a batch of images by adjusting brightness, contrast, saturation and hue

Parameters:

prob: float probability of layer to activate brightness: float upper limit of a random brightness shift. Lower is always 0 contrast: float, list of float shift of contrast. If a list of float is given, min value of that list serves as the lower boundary of contrast shift and the max value as the upper boundary. If a single float is given, lower and upper boundaries are determined as 1-contrast and 1+contrast saturation: float, list of float shift of saturation. If a list of float is given, min value of that list serves as the lower boundary of saturation shift and the second value as the upper boundary. If a single float is given, lower and upper boundaries are determined as 1-saturation and 1+saturation hue: float upper limit of random hue shift. Lower is always 0

Methods:

call: run the layer on given batch of images

Source code in conftrainer/augmentation/layers.py 
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
class RandomColorJitter(Augmentation):
    """
    Randomly Jitter a batch of images by adjusting brightness, contrast, saturation and hue

    Parameters:
    ----------
    prob: float
            probability of layer to activate
    brightness: float
        upper limit of a random brightness shift. Lower is always 0
    contrast: float, list of float
        shift of contrast. If a list of float is given, min value of that list serves as the
        lower boundary of contrast shift and the max value as the upper boundary. If a single
        float is given, lower and upper boundaries are determined as 1-contrast and 1+contrast
    saturation: float, list of float
        shift of saturation. If a list of float is given, min value of that list serves as the
        lower boundary of saturation shift and the second value as the upper boundary. If a single
        float is given, lower and upper boundaries are determined as 1-saturation and 1+saturation
    hue: float
        upper limit of random hue shift. Lower is always 0

    Methods:
    --------
    call: run the layer on given batch of images

    """

    def __init__(self, factor=0.9, brightness=0., contrast=0., saturation=0., hue=0., **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.max_brightness = brightness
        self.max_hue = hue
        if isinstance(contrast, list):
            self.lower_contrast = min(contrast)
            self.upper_contrast = max(contrast)
        else:
            self.lower_contrast, self.upper_contrast = 1 - contrast, 1 + contrast

        if isinstance(saturation, list):
            self.lower_saturation = min(saturation)
            self.upper_saturation = max(saturation)
        else:
            self.lower_saturation, self.upper_saturation = 1 - saturation, 1 + saturation

    def augment_single_image(self, image):
        """
        Randomly adjust the brightness, contrast, saturation and hue of an image
        """

        aug = tf.image.random_brightness(image, self.max_brightness)
        aug = tf.image.random_contrast(aug, self.lower_contrast, self.upper_contrast)
        aug = tf.image.random_saturation(aug, self.lower_saturation, self.upper_saturation)
        aug = tf.image.random_hue(aug, self.max_hue)

        return aug

augment_single_image(image)

Randomly adjust the brightness, contrast, saturation and hue of an image

Source code in conftrainer/augmentation/layers.py 
168
169
170
171
172
173
174
175
176
177
178
def augment_single_image(self, image):
    """
    Randomly adjust the brightness, contrast, saturation and hue of an image
    """

    aug = tf.image.random_brightness(image, self.max_brightness)
    aug = tf.image.random_contrast(aug, self.lower_contrast, self.upper_contrast)
    aug = tf.image.random_saturation(aug, self.lower_saturation, self.upper_saturation)
    aug = tf.image.random_hue(aug, self.max_hue)

    return aug

RandomZoomCrop

Bases: Augmentation

Randomly zoom and crop an image

Parameters:

prob: float the probability of layer to activate min_area: float minimal ratio of area of the original image to leave after cropping shape: tuple of int input/output shape of the images

Methods:

call: run the layer on a given batch of images

Source code in conftrainer/augmentation/layers.py 
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
225
226
227
228
229
230
231
232
233
234
235
class RandomZoomCrop(Augmentation):
    """
    Randomly zoom and crop an image

    Parameters:
    ----------
    prob: float
        the probability of layer to activate
    min_area: float
        minimal ratio of area of the original image to leave after cropping
    shape: tuple of int
        input/output shape of the images

    Methods:
    --------
    call: run the layer on a given batch of images
    """

    def __init__(self, min_area, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.min_area = min_area

    def augment_single_image(self, image):
        """
        Generate a random crop window, crop the image and resize it to given shape
        """

        orig_height, orig_width = self.shape[:2]
        area = tf.cast(orig_width * orig_height, dtype=tf.float32)
        target_area = tf.random.uniform([], self.min_area, 1.0, dtype=tf.float32) * area
        log_ratio = (tf.math.log(3 / 4), tf.math.log(4 / 3))
        aspect_ratio = tf.math.exp(
                tf.random.uniform([], *log_ratio, dtype=tf.float32))

        width = tf.cast(tf.round(tf.sqrt(target_area * aspect_ratio)), tf.int32)
        height = tf.cast(tf.round(tf.sqrt(target_area / aspect_ratio)), tf.int32)

        width = tf.minimum(width, orig_width)
        height = tf.minimum(height, orig_height)

        offset_w = tf.random.uniform((),
                                     minval=0,
                                     maxval=orig_width - width + 1,
                                     dtype=tf.int32)
        offset_h = tf.random.uniform((),
                                     minval=0,
                                     maxval=orig_height - height + 1,
                                     dtype=tf.int32)

        cropped = tf.image.crop_to_bounding_box(image=image, offset_height=offset_h,
                                                offset_width=offset_w,
                                                target_height=height, target_width=width)
        resized = tf.image.resize(cropped, self.shape[:2])
        resized = tf.image.convert_image_dtype(resized, self.compute_dtype)
        return resized

augment_single_image(image)

Generate a random crop window, crop the image and resize it to given shape

Source code in conftrainer/augmentation/layers.py 
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
def augment_single_image(self, image):
    """
    Generate a random crop window, crop the image and resize it to given shape
    """

    orig_height, orig_width = self.shape[:2]
    area = tf.cast(orig_width * orig_height, dtype=tf.float32)
    target_area = tf.random.uniform([], self.min_area, 1.0, dtype=tf.float32) * area
    log_ratio = (tf.math.log(3 / 4), tf.math.log(4 / 3))
    aspect_ratio = tf.math.exp(
            tf.random.uniform([], *log_ratio, dtype=tf.float32))

    width = tf.cast(tf.round(tf.sqrt(target_area * aspect_ratio)), tf.int32)
    height = tf.cast(tf.round(tf.sqrt(target_area / aspect_ratio)), tf.int32)

    width = tf.minimum(width, orig_width)
    height = tf.minimum(height, orig_height)

    offset_w = tf.random.uniform((),
                                 minval=0,
                                 maxval=orig_width - width + 1,
                                 dtype=tf.int32)
    offset_h = tf.random.uniform((),
                                 minval=0,
                                 maxval=orig_height - height + 1,
                                 dtype=tf.int32)

    cropped = tf.image.crop_to_bounding_box(image=image, offset_height=offset_h,
                                            offset_width=offset_w,
                                            target_height=height, target_width=width)
    resized = tf.image.resize(cropped, self.shape[:2])
    resized = tf.image.convert_image_dtype(resized, self.compute_dtype)
    return resized

RandomSolarize

Bases: Augmentation

Randomly solarize a batch of images, inverting pixels with low values.

Expects the inputs to be in range [0,255]

Parameters:

prob: float the probability of layer to activate threshold: int upper bound of pixels to invert

Methods:

call: run the layer on a given batch of images

Notes:

When paired with RandomInvert, the pixels of high values will be inverted instead

Source code in conftrainer/augmentation/layers.py 
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
class RandomSolarize(Augmentation):
    """
    Randomly solarize a batch of images, inverting pixels with low values.

    Expects the inputs to be in range [0,255]

    Parameters:
    ----------
    prob: float
        the probability of layer to activate
    threshold: int
        upper bound of pixels to invert

    Methods:
    --------
    call: run the layer on a given batch of images

    Notes:
    ------
    When paired with RandomInvert, the pixels of high values will be inverted instead
    """

    def __init__(self, factor, threshold=128, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.threshold = threshold

    def augment_single_image(self, image):
        """
        Solarize an image
        """
        return tf.where(image > self.threshold, image, 255 - image)

augment_single_image(image)

Solarize an image

Source code in conftrainer/augmentation/layers.py 
264
265
266
267
268
def augment_single_image(self, image):
    """
    Solarize an image
    """
    return tf.where(image > self.threshold, image, 255 - image)

RandomInvert

Bases: Augmentation

Randomly invert pixel values of a batch of images

Expects the inputs to be in range [0,255]

Parameters:

prob: float the probability of layer to activate

Methods:

call: run the layer on a given batch of images

Notes:

When paired with RandomSolarize, only the part of the pixels is inverted

Source code in conftrainer/augmentation/layers.py 
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
class RandomInvert(Augmentation):
    """
    Randomly invert pixel values of a batch of images

    Expects the inputs to be in range [0,255]

    Parameters:
    ----------
    prob: float
        the probability of layer to activate

    Methods:
    --------
    call: run the layer on a given batch of images

    Notes:
    ------
    When paired with RandomSolarize, only the part of the pixels is inverted
    """

    def __init__(self, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)

    def augment_single_image(self, image):
        """
        Invert pixel values of a single image
        """
        return tf.add(255., - image)

augment_single_image(image)

Invert pixel values of a single image

Source code in conftrainer/augmentation/layers.py 
294
295
296
297
298
def augment_single_image(self, image):
    """
    Invert pixel values of a single image
    """
    return tf.add(255., - image)

RandomBlur

Bases: Augmentation

Randomly blur a batch of images by applying a 2d Gaussian filter with a random sigma. For more info about the filter and its arguments, see tensorflow_addons.image.gaussian_filter2d

Parameters:

factor: float the probability of layer to activate filter_shape: tuple of int shape of the blur filter to apply. If tuple with more than 2 items is passed, only the first 2 number will be used sigma_low: float, default: 0.1 lower boundary of random sigma parameter to use when blurring sigma_high: float, default: 2 upper boundary of random sigma parameter to use when blurring

Methods:

call: run the layer on a given batch of images

Source code in conftrainer/augmentation/layers.py 
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
class RandomBlur(Augmentation):
    """
    Randomly blur a batch of images by applying a 2d Gaussian filter with a random sigma.
    For more info about the filter and its arguments, see tensorflow_addons.image.gaussian_filter2d

    Parameters:
    ----------
    factor: float
        the probability of layer to activate
    filter_shape: tuple of int
        shape of the blur filter to apply. If tuple with more than 2 items is passed, only the
        first 2 number will be used
    sigma_low: float, default: 0.1
        lower boundary of random sigma parameter to use when blurring
    sigma_high: float, default: 2
        upper boundary of random sigma parameter to use when blurring

    Methods:
    --------
    call: run the layer on a given batch of images
    """

    def __init__(self, factor, kernel_size, sigma_low=0.1, sigma_high=2, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.kernel_x, self.kernel_y = kernel_size[:2]
        self.sigma_low = sigma_low
        self.sigma_high = sigma_high

    def get_random_transformation(self):
        """Create gaussian kernels to augment the image"""
        sigma_x = tf.random.uniform((1,), minval=self.sigma_low, maxval=self.sigma_high)
        sigma_y = tf.random.uniform((1,), minval=self.sigma_low, maxval=self.sigma_high)
        blur_w = self.get_kernel(sigma_y, self.kernel_y)
        blur_h = self.get_kernel(sigma_x, self.kernel_x)
        blur_w = tf.reshape(blur_w, [self.kernel_y, 1, 1, 1])
        blur_h = tf.reshape(blur_h, [1, self.kernel_x, 1, 1])
        return blur_w, blur_h

    def augment_single_image(self, image):
        """Apply gaussian blurring to single image"""
        image = tf.expand_dims(image, axis=0)

        num_channels = tf.shape(image)[-1]
        blur_w, blur_h = self.get_random_transformation()
        blur_h = tf.cast(tf.tile(blur_h, [1, 1, num_channels, 1]), dtype=self.compute_dtype)
        blur_w = tf.cast(tf.tile(blur_w, [1, 1, num_channels, 1]), dtype=self.compute_dtype)

        blurred = tf.nn.depthwise_conv2d(image, blur_h, strides=[1, 1, 1, 1], padding="SAME")
        blurred = tf.nn.depthwise_conv2d(blurred, blur_w, strides=[1, 1, 1, 1], padding="SAME")

        return tf.squeeze(blurred, axis=0)

    @staticmethod
    def get_kernel(sigma, filter_size):
        """Create a gaussian kernel with given sigma and shape"""
        # Run in tf.float32 to avoid image corruption
        x = tf.cast(
                tf.range(-filter_size // 2 + 1, filter_size // 2 + 1),
                dtype=tf.float32,
        )
        blur_filter = tf.exp(
                -tf.pow(x, 2.0)
                / (2.0 * tf.pow(tf.cast(sigma, dtype=tf.float32), 2.0))
        )
        blur_filter /= tf.reduce_sum(blur_filter)
        return blur_filter

get_random_transformation()

Create gaussian kernels to augment the image

Source code in conftrainer/augmentation/layers.py 
329
330
331
332
333
334
335
336
337
def get_random_transformation(self):
    """Create gaussian kernels to augment the image"""
    sigma_x = tf.random.uniform((1,), minval=self.sigma_low, maxval=self.sigma_high)
    sigma_y = tf.random.uniform((1,), minval=self.sigma_low, maxval=self.sigma_high)
    blur_w = self.get_kernel(sigma_y, self.kernel_y)
    blur_h = self.get_kernel(sigma_x, self.kernel_x)
    blur_w = tf.reshape(blur_w, [self.kernel_y, 1, 1, 1])
    blur_h = tf.reshape(blur_h, [1, self.kernel_x, 1, 1])
    return blur_w, blur_h

augment_single_image(image)

Apply gaussian blurring to single image

Source code in conftrainer/augmentation/layers.py 
339
340
341
342
343
344
345
346
347
348
349
350
351
def augment_single_image(self, image):
    """Apply gaussian blurring to single image"""
    image = tf.expand_dims(image, axis=0)

    num_channels = tf.shape(image)[-1]
    blur_w, blur_h = self.get_random_transformation()
    blur_h = tf.cast(tf.tile(blur_h, [1, 1, num_channels, 1]), dtype=self.compute_dtype)
    blur_w = tf.cast(tf.tile(blur_w, [1, 1, num_channels, 1]), dtype=self.compute_dtype)

    blurred = tf.nn.depthwise_conv2d(image, blur_h, strides=[1, 1, 1, 1], padding="SAME")
    blurred = tf.nn.depthwise_conv2d(blurred, blur_w, strides=[1, 1, 1, 1], padding="SAME")

    return tf.squeeze(blurred, axis=0)

get_kernel(sigma, filter_size) staticmethod

Create a gaussian kernel with given sigma and shape

Source code in conftrainer/augmentation/layers.py 
353
354
355
356
357
358
359
360
361
362
363
364
365
366
@staticmethod
def get_kernel(sigma, filter_size):
    """Create a gaussian kernel with given sigma and shape"""
    # Run in tf.float32 to avoid image corruption
    x = tf.cast(
            tf.range(-filter_size // 2 + 1, filter_size // 2 + 1),
            dtype=tf.float32,
    )
    blur_filter = tf.exp(
            -tf.pow(x, 2.0)
            / (2.0 * tf.pow(tf.cast(sigma, dtype=tf.float32), 2.0))
    )
    blur_filter /= tf.reduce_sum(blur_filter)
    return blur_filter

RandomGrayscale

Bases: Augmentation

Randomly turn a batch of images to Grayscale

Parameters:

prob: float the probability of layer to activate

Methods:

call: run the layer on a given batch of images

Source code in conftrainer/augmentation/layers.py 
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
class RandomGrayscale(Augmentation):
    """
    Randomly turn a batch of images to Grayscale

    Parameters:
    ----------
    prob: float
        the probability of layer to activate

    Methods:
    --------
    call: run the layer on a given batch of images
    """

    def __init__(self, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.rgb_weights = RGB_WEIGHTS

    def augment_single_image(self, image):
        """
        Turn an image to grayscale
        """
        gray = tf.reduce_sum(image * self.rgb_weights, axis=-1)
        gray_3d = tf.stack([gray, gray, gray], axis=-1)

        return gray_3d

augment_single_image(image)

Turn an image to grayscale

Source code in conftrainer/augmentation/layers.py 
387
388
389
390
391
392
393
394
def augment_single_image(self, image):
    """
    Turn an image to grayscale
    """
    gray = tf.reduce_sum(image * self.rgb_weights, axis=-1)
    gray_3d = tf.stack([gray, gray, gray], axis=-1)

    return gray_3d

RandomSepia

Bases: Augmentation

Randomly turn images to sepia color scheme

Parameters:

Name Type Description Default
factor

probability of the layer to activate

 required
Source code in conftrainer/augmentation/layers.py 
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
class RandomSepia(Augmentation):
    """
    Randomly turn images to sepia color scheme

    Parameters
    ----------
    factor: float
        probability of the layer to activate
    """

    def __init__(self, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.red, self.green, self.blue = SEPIA_RGB_WEIGHTS

    def augment_single_image(self, image):
        """
        Turn a single image to sepia
        """
        red = tf.math.reduce_sum(image * self.red, axis=-1)
        blue = tf.math.reduce_sum(image * self.blue, axis=-1)
        green = tf.math.reduce_sum(image * self.green, axis=-1)
        return tf.stack([red, green, blue], axis=-1)

augment_single_image(image)

Turn a single image to sepia

Source code in conftrainer/augmentation/layers.py 
411
412
413
414
415
416
417
418
def augment_single_image(self, image):
    """
    Turn a single image to sepia
    """
    red = tf.math.reduce_sum(image * self.red, axis=-1)
    blue = tf.math.reduce_sum(image * self.blue, axis=-1)
    green = tf.math.reduce_sum(image * self.green, axis=-1)
    return tf.stack([red, green, blue], axis=-1)

RandomChannelShuffle

Bases: Augmentation

Randomly Shuffle the order of channels of RGB images

Parameters:

Name Type Description Default
factor

probability of layer to activate on each call

 required
Source code in conftrainer/augmentation/layers.py 
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
class RandomChannelShuffle(Augmentation):
    """
    Randomly Shuffle the order of channels of RGB images

    Parameters
    ----------
    factor: float
        probability of layer to activate on each call
    """

    def __init__(self, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.channel_order = [0, 1, 2]

    def augment_single_image(self, image):
        """
        Shuffle the order of channels of a single image
        """
        channel_order = tf.random.shuffle(self.channel_order)
        return tf.gather(params=image, indices=channel_order, axis=-1)

augment_single_image(image)

Shuffle the order of channels of a single image

Source code in conftrainer/augmentation/layers.py 
435
436
437
438
439
440
def augment_single_image(self, image):
    """
    Shuffle the order of channels of a single image
    """
    channel_order = tf.random.shuffle(self.channel_order)
    return tf.gather(params=image, indices=channel_order, axis=-1)

RandomChannelShift

Bases: Augmentation

Randomly shift one or multiple channel's of RGB images. Expects the pixel values to be in range [0, 255]

Parameters:

Name Type Description Default
factor

probability of the layer to activate

 required
channels

list of indices of channels to shift

 None
delta

maximal magnitude of distortions

 0
Source code in conftrainer/augmentation/layers.py 
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
class RandomChannelShift(Augmentation):
    """
    Randomly shift one or multiple channel's of RGB images. Expects the pixel values to be in
    range [0, 255]

    Parameters
    ----------
    factor: float
        probability of the layer to activate
    channels: list of int
        list of indices of channels to shift
    delta: int
        maximal magnitude of distortions
    """

    def __init__(self, factor, channels=None, delta=0, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.channels = [0, 0, 0]
        if channels:
            for ind in channels:
                self.channels[ind] = 1
        self.delta = delta

    def augment_single_image(self, image):
        """
        Shift pixel values of given channels of a single image by random values
        """
        shift_per_channel = tf.random.uniform(shape=(1, 3),
                                              minval=-self.delta,
                                              maxval=self.delta)
        shift = shift_per_channel * self.channels

        return tf.clip_by_value(tf.add(image, shift), 0, 255)

augment_single_image(image)

Shift pixel values of given channels of a single image by random values

Source code in conftrainer/augmentation/layers.py 
466
467
468
469
470
471
472
473
474
475
def augment_single_image(self, image):
    """
    Shift pixel values of given channels of a single image by random values
    """
    shift_per_channel = tf.random.uniform(shape=(1, 3),
                                          minval=-self.delta,
                                          maxval=self.delta)
    shift = shift_per_channel * self.channels

    return tf.clip_by_value(tf.add(image, shift), 0, 255)

RandomAdjustJpegQuality

Bases: Augmentation

Randomly compress the Jpeg quality of a batch of images.

The pixel values of images are expected to be in range [0, 255]

Parameters:

Name Type Description Default
factor

probability of the layer to activate

 required
min_quality

positive integer < 100, minimal quality of image ofter distortion

 25
Source code in conftrainer/augmentation/layers.py 
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
class RandomAdjustJpegQuality(Augmentation):
    """
    Randomly compress the Jpeg quality of a batch of images.

    The pixel values of images are expected to be in range [0, 255]

    Parameters
    ----------
    factor: float
        probability of the layer to activate
    min_quality: int
        positive integer < 100, minimal quality of image ofter distortion
    """

    def __init__(self, factor, min_quality=25, max_quality=100, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.min_quality = min_quality
        self.max_quality = max_quality

    def augment_single_image(self, image):
        """
        Randomly adjust jpeg quality of a single image
        """
        return 255 * tf.image.random_jpeg_quality(image / 255,
                                                  min_jpeg_quality=self.min_quality,
                                                  max_jpeg_quality=self.max_quality)

augment_single_image(image)

Randomly adjust jpeg quality of a single image

Source code in conftrainer/augmentation/layers.py 
497
498
499
500
501
502
503
def augment_single_image(self, image):
    """
    Randomly adjust jpeg quality of a single image
    """
    return 255 * tf.image.random_jpeg_quality(image / 255,
                                              min_jpeg_quality=self.min_quality,
                                              max_jpeg_quality=self.max_quality)

RandomPixelDropout

Bases: Augmentation

Randomly turn pixel values of an image to 0

Parameters:

Name Type Description Default
factor

probability of the layer to activate

 required
dropout_probability

probability of each pixel value to be turned to 0

 0.001
Source code in conftrainer/augmentation/layers.py 
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
class RandomPixelDropout(Augmentation):
    """
    Randomly turn pixel values of an image to 0

    Parameters
    ----------
    factor: float
        probability of the layer to activate
    dropout_probability: float
        probability of each pixel value to be turned to 0
    """

    def __init__(self, factor, dropout_probability=0.001, **kwargs):
        super().__init__(factor=factor, **kwargs)
        self.dropout_probability = dropout_probability

    def augment_single_image(self, image):
        """
        Randomly dropout pixel values of a single image
        """

        mask = tf.random.uniform(shape=tf.shape(image)[:2], minval=0, maxval=1)
        mask = tf.where(mask > self.dropout_probability, 1., 0.)
        mask = tf.stack([mask, mask, mask], axis=-1)
        return image * mask

augment_single_image(image)

Randomly dropout pixel values of a single image

Source code in conftrainer/augmentation/layers.py 
522
523
524
525
526
527
528
529
530
def augment_single_image(self, image):
    """
    Randomly dropout pixel values of a single image
    """

    mask = tf.random.uniform(shape=tf.shape(image)[:2], minval=0, maxval=1)
    mask = tf.where(mask > self.dropout_probability, 1., 0.)
    mask = tf.stack([mask, mask, mask], axis=-1)
    return image * mask

RandomHorizontalFlip

Bases: Augmentation

Randomly flip images horizontally

Source code in conftrainer/augmentation/layers.py 
533
534
535
536
537
538
539
540
541
542
543
544
545
class RandomHorizontalFlip(Augmentation):
    """
    Randomly flip images horizontally
    """

    def __init__(self, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)

    def augment_single_image(self, image):
        """
        Flip single image
        """
        return tf.image.flip_left_right(image)

augment_single_image(image)

Flip single image

Source code in conftrainer/augmentation/layers.py 
541
542
543
544
545
def augment_single_image(self, image):
    """
    Flip single image
    """
    return tf.image.flip_left_right(image)

RandomVerticalFlip

Bases: Augmentation

Randomly flip images vertically

Source code in conftrainer/augmentation/layers.py 
548
549
550
551
552
553
554
555
556
557
558
559
560
class RandomVerticalFlip(Augmentation):
    """
    Randomly flip images vertically
    """

    def __init__(self, factor, **kwargs):
        super().__init__(factor=factor, **kwargs)

    def augment_single_image(self, image):
        """
        Flip single image
        """
        return tf.image.flip_up_down(image)

augment_single_image(image)

Flip single image

Source code in conftrainer/augmentation/layers.py 
556
557
558
559
560
def augment_single_image(self, image):
    """
    Flip single image
    """
    return tf.image.flip_up_down(image)