Skip to content

random_cutout

Random Cutout Layer API

This module provides classes to build random cutout layers.

Classes:

Classes

RandomCutout1D 

RandomCutout1D(factor: float | tuple[float, float] = 0.1, cutouts: int = 1, fill_mode='constant', fill_value: float = 0.0, **kwargs)

Apply random cutout to the input. This is similar to its 2D counterpart where a random portion of the input is cutout. We allow providing a range for the factor and cutouts to randomly pick the values.

Parameters:

  • factor (float | tuple[float, float], default: 0.1 ) –

    Factor of the duration to cutout. If tuple, factor is randomly picked between the values.

  • cutouts (int, default: 1 ) –

    Number of cutouts to apply.

  • fill_mode (str, default: 'constant' ) –

    Fill mode. "constant" or "normal".

  • fill_value (float, default: 0.0 ) –

    Fill value for the cutout.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
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
def __init__(
    self,
    factor: float | tuple[float, float] = 0.1,
    cutouts: int = 1,
    fill_mode="constant",
    fill_value: float = 0.0,
    **kwargs,
):
    """Apply random cutout to the input. This is similar to its 2D counterpart where a random portion of the input is cutout.
    We allow providing a range for the factor and cutouts to randomly pick the values.

    Args:
        factor (float|tuple[float,float]): Factor of the duration to cutout. If tuple, factor is randomly picked between the values.
        cutouts (int): Number of cutouts to apply.
        fill_mode (str): Fill mode. "constant" or "normal".
        fill_value (float): Fill value for the cutout.

    """
    super().__init__(**kwargs)

    self.factor = parse_factor(factor, min_value=0, max_value=1, param_name="factor")
    self.cutouts = cutouts
    self.fill_mode = fill_mode
    self.fill_value = fill_value
    if fill_mode not in ["normal", "constant"]:
        raise ValueError(f'`fill_mode` should be "normal" or "constant". Got `fill_mode`={fill_mode}')

Functions

call 
call(inputs, training: bool = True)

Override the call method to apply multiple cutouts.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
54
55
56
57
58
59
60
61
62
63
64
65
66
def call(self, inputs, training: bool = True):
    """Override the call method to apply multiple cutouts."""
    self.training = training
    if not self.training:
        return inputs
    inputs, metadata = self._format_inputs(inputs)
    outputs = keras.ops.fori_loop(
        lower=0,
        upper=self.cutouts,
        body_fun=lambda _, x: self.batch_augment(x),
        init_val=self.batch_augment(inputs),
    )
    return self._format_outputs(outputs, metadata)
get_random_transformations 
get_random_transformations(input_shape)

Generate random cutout locations, sizes, and fill values.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
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
def get_random_transformations(self, input_shape):
    """Generate random cutout locations, sizes, and fill values."""
    batch_size = input_shape[0]
    duration_size = input_shape[self.data_axis]

    cut_size = keras.random.randint(
        shape=(batch_size,),
        minval=int(duration_size * self.factor[0]),
        maxval=int(duration_size * self.factor[1]) + 1,
        dtype="int32",
        seed=self.random_generator,
    )
    cut_start = keras.random.randint(
        shape=(batch_size,),
        minval=0,
        maxval=int(duration_size * (1 - self.factor[1]) + 1),
        dtype="int32",
        seed=self.random_generator,
    )
    if self.fill_mode == "constant":
        fill = keras.ops.ones(input_shape) * self.fill_value
    else:
        fill = keras.random.normal(input_shape, mean=0, stddev=self.fill_value, seed=self.random_generator)

    return {
        "cut_start": cut_start,
        "cut_size": cut_size,
        "fill": fill,
    }
augment_sample 
augment_sample(inputs) -> keras.KerasTensor

Apply cutout to the input.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
 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
def augment_sample(self, inputs) -> keras.KerasTensor:
    """Apply cutout to the input."""
    sample = inputs[self.SAMPLES]
    transforms = inputs[self.TRANSFORMS]
    cut_start = transforms["cut_start"]
    cut_size = transforms["cut_size"]
    fill = transforms["fill"]

    duration_size = sample.shape[self.data_axis]
    ch_size = sample.shape[self.ch_axis]

    if self.data_format == "channels_first":
        reshape_size = (1, duration_size)
        tile_size = (ch_size, 1)
    else:
        reshape_size = (duration_size, 1)
        tile_size = (1, ch_size)

    mask = keras.ops.tile(
        keras.ops.reshape(
            keras.ops.logical_and(
                keras.ops.arange(duration_size) >= cut_start, keras.ops.arange(duration_size) < cut_start + cut_size
            ),
            reshape_size,
        ),
        tile_size,
    )
    result = keras.ops.where(mask, fill, sample)
    return result
get_config 
get_config()

Serialize the configuration of the layer.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
128
129
130
131
132
133
134
135
136
137
138
139
def get_config(self):
    """Serialize the configuration of the layer."""
    config = super().get_config()
    config.update(
        {
            "factor": self.factor,
            "cutouts": self.cutouts,
            "fill_mode": self.fill_mode,
            "fill_value": self.fill_value,
        }
    )
    return config

RandomCutout2D 

RandomCutout2D(factor: float | tuple[float, float] = 0.1, cutouts: int = 1, fill_mode='constant', fill_value: float = 0.0, **kwargs)

Apply random cutout to the input. This is similar to its 1D counterpart where a random portion of the input is cutout. We allow providing a range for the factor and cutouts to randomly pick the values.

Parameters:

  • factor (float | tuple[float, float], default: 0.1 ) –

    Factor of the dimensions to cutout. If tuple, factor is randomly picked between the values.

  • cutouts (int, default: 1 ) –

    Number of cutouts to apply.

  • fill_mode (str, default: 'constant' ) –

    Fill mode. "constant" or "normal".

  • fill_value (float, default: 0.0 ) –

    Fill value for the cutout.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
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
def __init__(
    self,
    factor: float | tuple[float, float] = 0.1,
    cutouts: int = 1,
    fill_mode="constant",
    fill_value: float = 0.0,
    **kwargs,
):
    """Apply random cutout to the input. This is similar to its 1D counterpart where a random portion of the input is cutout.
    We allow providing a range for the factor and cutouts to randomly pick the values.

    Args:
        factor (float|tuple[float,float]): Factor of the dimensions to cutout. If tuple, factor is randomly picked between the values.
        cutouts (int): Number of cutouts to apply.
        fill_mode (str): Fill mode. "constant" or "normal".
        fill_value (float): Fill value for the cutout.

    """
    super().__init__(**kwargs)

    self.factor = parse_factor(factor, min_value=0, max_value=1, param_name="factor")
    self.cutouts = cutouts
    self.fill_mode = fill_mode
    self.fill_value = fill_value
    if fill_mode not in ["normal", "constant"]:
        raise ValueError(f'`fill_mode` should be "normal" or "constant". Got `fill_mode`={fill_mode}')

Functions

call 
call(inputs, training: bool = True)

Override the call method to apply multiple cutouts.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
177
178
179
180
181
182
183
184
185
186
187
188
189
def call(self, inputs, training: bool = True):
    """Override the call method to apply multiple cutouts."""
    self.training = training
    if not self.training:
        return inputs
    inputs, metadata = self._format_inputs(inputs)
    outputs = keras.ops.fori_loop(
        lower=0,
        upper=self.cutouts,
        body_fun=lambda _, x: self.batch_augment(x),
        init_val=self.batch_augment(inputs),
    )
    return self._format_outputs(outputs, metadata)
get_random_transformations 
get_random_transformations(input_shape)

Generate random cutout locations, sizes, and fill values.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
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
236
def get_random_transformations(self, input_shape):
    """Generate random cutout locations, sizes, and fill values."""
    batch_size = input_shape[0]
    height_size = input_shape[self.height_axis]
    width_size = input_shape[self.width_axis]

    cut_height = keras.random.randint(
        shape=(batch_size,),
        minval=int(height_size * self.factor[0]),
        maxval=int(height_size * self.factor[1]) + 1,
        dtype="int32",
        seed=self.random_generator,
    )
    cut_width = keras.random.randint(
        shape=(batch_size,),
        minval=int(width_size * self.factor[0]),
        maxval=int(width_size * self.factor[1]) + 1,
        dtype="int32",
        seed=self.random_generator,
    )
    cut_start_height = keras.random.randint(
        shape=(batch_size,),
        minval=0,
        maxval=int(height_size * (1 - self.factor[1]) + 1),
        dtype="int32",
        seed=self.random_generator,
    )
    cut_start_width = keras.random.randint(
        shape=(batch_size,),
        minval=0,
        maxval=int(width_size * (1 - self.factor[1]) + 1),
        dtype="int32",
        seed=self.random_generator,
    )
    if self.fill_mode == "constant":
        fill = keras.ops.ones(input_shape) * self.fill_value
    else:
        fill = keras.random.normal(input_shape, mean=0, stddev=self.fill_value, seed=self.random_generator)

    return {
        "cut_start_height": cut_start_height,
        "cut_start_width": cut_start_width,
        "cut_height": cut_height,
        "cut_width": cut_width,
        "fill": fill,
    }
augment_sample 
augment_sample(inputs) -> keras.KerasTensor

Apply cutout to the input.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.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
269
270
271
272
273
274
275
276
def augment_sample(self, inputs) -> keras.KerasTensor:
    """Apply cutout to the input."""
    sample = inputs[self.SAMPLES]
    transforms = inputs[self.TRANSFORMS]
    cut_start_height = transforms["cut_start_height"]
    cut_start_width = transforms["cut_start_width"]
    cut_height = transforms["cut_height"]
    cut_width = transforms["cut_width"]
    fill = transforms["fill"]

    height_size = sample.shape[self.height_axis]
    width_size = sample.shape[self.width_axis]
    ch_size = sample.shape[self.ch_axis]

    if self.data_format == "channels_first":
        reshape_size = (1, height_size, width_size)
        tile_size = (ch_size, 1, 1)
    else:
        reshape_size = (height_size, width_size, 1)
        tile_size = (1, 1, ch_size)

    mask = keras.ops.tile(
        keras.ops.reshape(
            keras.ops.logical_and(
                keras.ops.logical_and(
                    keras.ops.arange(height_size) >= cut_start_height,
                    keras.ops.arange(height_size) < cut_start_height + cut_height,
                ),
                keras.ops.logical_and(
                    keras.ops.arange(width_size) >= cut_start_width,
                    keras.ops.arange(width_size) < cut_start_width + cut_width,
                ),
            ),
            reshape_size,
        ),
        tile_size,
    )
    result = keras.ops.where(mask, fill, sample)
    return result
get_config 
get_config()

Serialize the configuration of the layer.

Source code in neuralspot_edge/layers/preprocessing/random_cutout.py 
278
279
280
281
282
283
284
285
286
287
288
289
def get_config(self):
    """Serialize the configuration of the layer."""
    config = super().get_config()
    config.update(
        {
            "factor": self.factor,
            "cutouts": self.cutouts,
            "fill_mode": self.fill_mode,
            "fill_value": self.fill_value,
        }
    )
    return config