Skip to content

mobileone

MobileOne

Overview

MobileOne is a fully convolutional neural network designed to have minimal latency when running in mobile/edge devices. The architecture consists of a series of depthwise separable convolutions and squeeze and excitation (SE) blocks. The network also uses standard batch normalization and ReLU activations that can be easily fused into the convolutional layers. Lastly, the network uses over-parameterized branches to improve training, yet can be merged into a single branch during inference.

For more info, refer to the original paper MobileOne: An Improved One millisecond Mobile Backbone.

Classes:

Functions:

Additions

The MobileOne architecture has been modified to allow the following:

  • Enable 1D and 2D variants.
  • Enable dilated convolutions.

Usage

import keras
from neuralspot_edge.models import MobileOne, MobileOneParams, MobileOneBlockParams

inputs = keras.Input(shape=(800, 1), name="inputs")

model = MobileOne.model_from_params(
    x=inputs,
    params=MobileOneParams(
        input_filters=24,
        input_kernel_size=(1, 7),
        input_strides=(1, 2),
        blocks=[
            MobileOneBlockParams(filters=32, depth=2, kernel_size=(1, 7), strides=(1, 2), se_ratio=2, se_depth=2, num_conv_branches=2)
        ],
        include_top=True,
        model_name="MobileOne",
    ),
)

model.summary()

Classes

MobileOneBlockParams

MobileOne block parameters

Attributes:

  • filters (int) –

    Number of filters

  • depth (int) –

    Layer depth

  • kernel_size (int | tuple[int, int]) –

    Kernel size

  • strides (int | tuple[int, int]) –

    Stride size

  • padding (int | tuple[int, int]) –

    Padding size

  • se_ratio (float) –

    Squeeze Excite ratio

  • se_depth (int) –

    Depth length to apply SE

  • num_conv_branches (int) –

    Number of conv branches

  • activation (str) –

    Activation function

MobileOneParams

MobileOne parameters

Attributes:

MobileOneModel

Helper class to generate model from parameters

Functions

layer_from_params staticmethod 
layer_from_params(inputs: keras.Input, params: MobileOneParams | dict, num_classes: int | None = None)

Create layer from parameters

Source code in neuralspot_edge/models/mobileone.py 
346
347
348
349
350
351
@staticmethod
def layer_from_params(inputs: keras.Input, params: MobileOneParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = MobileOneParams(**params)
    return mobileone_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: MobileOneParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/mobileone.py 
353
354
355
356
357
@staticmethod
def model_from_params(inputs: keras.Input, params: MobileOneParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = MobileOneModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

mobileone_block 

mobileone_block(output_filters: int, kernel_size: int | tuple[int, int] = 3, strides: int | tuple[int, int] = 1, padding: int | tuple[int, int] = 0, groups: int = 1, dilation: int = 1, inference_mode: bool = False, se_ratio: int = 0, num_conv_branches: int = 1, activation: str = 'relu6', name: str | None = None) -> keras.Layer

MBConv block w/ expansion and SE

Parameters:

  • output_filters (int) –

    Number of output filter channels

  • kernel_size (int | tuple[int, int], default: 3 ) –

    Kernel size. Defaults to 3.

  • strides (int | tuple[int, int], default: 1 ) –

    Stride length. Defaults to 1.

  • se_ratio (float, default: 0 ) –

    SE ratio. Defaults to 8.

  • name (str | None, default: None ) –

    Block name. Defaults to None.

Returns:

  • Layer –

    keras.Layer: Functional layer

Source code in neuralspot_edge/models/mobileone.py 
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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
def mobileone_block(
    output_filters: int,
    kernel_size: int | tuple[int, int] = 3,
    strides: int | tuple[int, int] = 1,
    padding: int | tuple[int, int] = 0,
    groups: int = 1,
    dilation: int = 1,
    inference_mode: bool = False,
    se_ratio: int = 0,
    num_conv_branches: int = 1,
    activation: str = "relu6",
    name: str | None = None,
) -> keras.Layer:
    """MBConv block w/ expansion and SE

    Args:
        output_filters (int): Number of output filter channels
        kernel_size (int | tuple[int, int], optional): Kernel size. Defaults to 3.
        strides (int | tuple[int, int], optional): Stride length. Defaults to 1.
        se_ratio (float, optional): SE ratio. Defaults to 8.
        name (str|None, optional): Block name. Defaults to None.

    Returns:
        keras.Layer: Functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        input_filters = x.shape[-1]
        stride_len = strides if isinstance(strides, int) else sum(strides) / len(strides)
        kernel_len = kernel_size if isinstance(kernel_size, int) else sum(kernel_size) / len(kernel_size)
        is_downsample = stride_len > 1
        is_depthwise = groups > 1 and groups == input_filters
        has_skip_branch = output_filters == input_filters and stride_len == 1

        if inference_mode:
            y = keras.layers.ZeroPadding2D(padding=padding)(x)
            y = conv2d(
                output_filters,
                kernel_size=kernel_size,
                strides=strides,
                padding="valid",
                dilation=dilation,
                groups=groups,
                use_bias=True,
                name=name,
            )(y)
            if se_ratio > 0:
                name_se = f"{name}.se" if name else None
                y = se_layer(ratio=se_ratio, name=name_se)(y)
            # END IF
            y = keras.layers.Activation(activation, name=f"{name}.act" if name else None)(y)
            return y
        # END IF

        branches = []

        # Skip branch
        if has_skip_branch:
            name_skip = f"{name}.skip" if name else None
            y_skip = batch_normalization(name=name_skip)(x)
            branches.append(y_skip)
        # END IF

        # Either groups is input_filters or is 1

        # Scale branch
        if kernel_len > 1:
            name_scale = f"{name}.scale" if name else None
            if is_depthwise:
                y_scale = keras.layers.DepthwiseConv2D(
                    kernel_size=(1, 1),
                    strides=(1, 1),  # strides,
                    padding="valid",
                    use_bias=False,
                    depthwise_initializer="he_normal",
                    name=f"{name_scale}.conv" if name_scale else None,
                )(x)
                y_scale = batch_normalization(name=name_scale)(y_scale)
                if is_downsample:
                    y_scale = keras.layers.MaxPool2D(pool_size=strides, padding="same")(y_scale)
            else:
                y_scale = keras.layers.Conv2D(
                    output_filters,
                    kernel_size=(1, 1),
                    strides=strides,
                    padding="valid",
                    groups=groups,
                    use_bias=False,
                    kernel_initializer="he_normal",
                    name=f"{name_scale}.conv" if name_scale else None,
                )(x)
                y_scale = batch_normalization(name=name_scale)(y_scale)
            branches.append(y_scale)
        # END IF

        # Other branches
        yp = keras.layers.ZeroPadding2D(padding=padding)(x)
        for b in range(num_conv_branches):
            name_branch = f"{name}.branch{b+1}" if name else None
            if is_depthwise:
                y_branch = keras.layers.DepthwiseConv2D(
                    kernel_size=kernel_size,
                    strides=(1, 1),
                    padding="valid",
                    use_bias=False,
                    depthwise_initializer="he_normal",
                    name=f"{name_branch}.conv" if name else None,
                )(yp)
                y_branch = batch_normalization(name=name_branch)(y_branch)
                if is_downsample:
                    y_branch = keras.layers.MaxPool2D(pool_size=strides, padding="same")(y_branch)
            else:
                y_branch = keras.layers.Conv2D(
                    output_filters,
                    kernel_size=kernel_size,
                    strides=strides,
                    groups=groups,
                    padding="valid",
                    use_bias=False,
                    kernel_initializer="he_normal",
                    name=f"{name_branch}.conv" if name else None,
                )(yp)
                y_branch = batch_normalization(name=name_branch)(y_branch)
            branches.append(y_branch)
        # END FOR

        # Merge branches
        y = keras.layers.Add(name=f"{name}.add" if name else None)(branches)

        # Squeeze-Excite block
        if se_ratio > 0:
            name_se = f"{name}.se" if name else None
            y = se_layer(ratio=se_ratio, name=name_se)(y)
        # END IF
        y = keras.layers.Activation(activation, name=f"{name}.act" if name else None)(y)
        return y

    # END DEF
    return layer

mobileone_layer 

mobileone_layer(x: keras.KerasTensor, params: MobileOneParams, num_classes: int | None = None, inference_mode: bool = False) -> keras.KerasTensor

Create MobileOne TF functional model

Parameters:

  • x (KerasTensor) –

    Input tensor

  • params (MobileOneParams) –

    Model parameters.

  • num_classes (int, default: None ) –

    Number of classes.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/mobileone.py 
262
263
264
265
266
267
268
269
270
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
299
300
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
def mobileone_layer(
    x: keras.KerasTensor,
    params: MobileOneParams,
    num_classes: int | None = None,
    inference_mode: bool = False,
) -> keras.KerasTensor:
    """Create MobileOne TF functional model

    Args:
        x (keras.KerasTensor): Input tensor
        params (MobileOneParams): Model parameters.
        num_classes (int, optional): Number of classes.

    Returns:
        keras.KerasTensor: Output tensor
    """

    requires_reshape = len(x.shape) == 3
    if requires_reshape:
        y = keras.layers.Reshape((1,) + x.shape[1:])(x)
    else:
        y = x
    # END IF

    y = mobileone_block(
        output_filters=params.input_filters,
        kernel_size=params.input_kernel_size,
        strides=params.input_strides,
        padding=params.input_padding,
        groups=1,
        inference_mode=inference_mode,
        name=f"M0.B{0}.D{0}.DW",
    )(y)

    for b, block in enumerate(params.blocks):
        for d in range(block.depth):
            se_ratio = block.se_ratio if d >= block.depth - block.se_depth else 0
            # Depthwise block
            y = mobileone_block(
                output_filters=y.shape[-1],
                kernel_size=block.kernel_size,
                strides=block.strides if d == 0 else (1, 1),
                padding=block.padding,
                groups=y.shape[-1],
                inference_mode=inference_mode,
                se_ratio=se_ratio,
                num_conv_branches=block.num_conv_branches,
                name=f"M1.B{b+1}.D{d+1}.DW",
            )(y)

            # Pointwise block
            y = mobileone_block(
                output_filters=block.filters,
                kernel_size=(1, 1),
                strides=(1, 1),
                padding=(0, 0),
                groups=1,
                inference_mode=inference_mode,
                se_ratio=se_ratio,
                num_conv_branches=block.num_conv_branches,
                name=f"M1.B{b+1}.D{d+1}.PW",
            )(y)
        # END FOR
    # END FOR

    if params.include_top:
        name = "top"
        y = keras.layers.GlobalAveragePooling2D(name=f"{name}.pool")(y)
        if 0 < params.dropout < 1:
            y = keras.layers.Dropout(params.dropout)(y)
        if num_classes is not None:
            y = keras.layers.Dense(num_classes, name=name)(y)
        if params.output_activation:
            y = keras.layers.Activation(params.output_activation)(y)

    if requires_reshape:
        y = keras.layers.Reshape(y.shape[2:])(y)

    return y