Skip to content

regnet

RegNet Model API

This module provides utility functions to generate RegNet models.

Fore more information, please refer to the following paper: https://arxiv.org/abs/2101.00590

Classes:

Functions:

Classes

RegNetBlockParam

RegNet block parameters

Attributes:

  • filters (int) –

    Number of filters

  • depth (int) –

    Layer depth

  • group_width (int) –

    Group width

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

    Kernel size

  • strides (int | tuple[int, int]) –

    Stride size

  • se_ratio (float) –

    Squeeze Excite ratio

  • droprate (float) –

    Drop rate

  • activation (str) –

    Activation function

RegNetParams

RegNet parameters

Attributes:

blocks (list[RegNetBlockParam]): RegNet blocks
input_filters (int): Input filters
input_strides (int | tuple[int, int]): Input stride
input_activation (str): Input activation
output_filters (int): Output filters
block_style (Literal["y", "z"]): Block style
include_top (bool): Include top
output_activation (str | None): Output activation
dropout (float): Dropout rate
name (str): Model name

RegNetModel

Helper class to generate model from parameters

Functions

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

Create layer from parameters

Source code in neuralspot_edge/models/regnet.py 
310
311
312
313
314
315
@staticmethod
def layer_from_params(inputs: keras.Input, params: RegNetParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = RegNetParams(**params)
    return regnet_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: RegNetParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/regnet.py 
317
318
319
320
321
@staticmethod
def model_from_params(inputs: keras.Input, params: RegNetParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = RegNetModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

yblock 

yblock(output_filters: int = 0, group_width: int = 0, kernel_size: int | tuple[int, int] = 3, strides: int | tuple[int, int] = 1, se_ratio: float = 4, activation: str = 'relu6', name: str | None = None) -> Callable[[keras.KerasTensor], keras.KerasTensor]

RegNet Y-Block

Parameters:

  • output_filters (int, default: 0 ) –

    Number of output filters. Defaults to 0.

  • group_width (int, default: 0 ) –

    Group width. Defaults to 0.

  • 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: 4 ) –

    SE ratio. Defaults to 4.

  • activation (str, default: 'relu6' ) –

    Activation function. Defaults to "relu6".

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

    Block name. Defaults to None.

Returns:

  • Callable[[KerasTensor], KerasTensor] –

    Callable[[keras.KerasTensor], keras.KerasTensor]: Functional layer

Source code in neuralspot_edge/models/regnet.py 
 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
def yblock(
    output_filters: int = 0,
    group_width: int = 0,
    kernel_size: int | tuple[int, int] = 3,
    strides: int | tuple[int, int] = 1,
    se_ratio: float = 4,
    activation: str = "relu6",
    name: str | None = None,
) -> Callable[[keras.KerasTensor], keras.KerasTensor]:
    """RegNet Y-Block

    Args:
        output_filters (int, optional): Number of output filters. Defaults to 0.
        group_width (int, optional): Group width. Defaults to 0.
        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 4.
        activation (str, optional): Activation function. Defaults to "relu6".
        name (str|None, optional): Block name. Defaults to None.

    Returns:
        Callable[[keras.KerasTensor], keras.KerasTensor]: Functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        input_filters = x.shape[-1]
        groups = output_filters // group_width
        use_skip = (input_filters != output_filters) or (strides != 1 if isinstance(strides, int) else strides[0] != 1)

        name_ex = f"{name}.exp" if name else None
        y = conv2d(output_filters, kernel_size=(1, 1), name=name_ex)(x)
        y = batch_normalization(name=name_ex)(y)
        y = keras.layers.Activation(activation, name=f"{name_ex}.act")(y)

        y = keras.layers.Conv2D(
            output_filters,
            kernel_size=kernel_size,
            strides=strides,
            groups=groups,
            padding="same",
            use_bias=False,
            kernel_initializer=keras.initializers.VarianceScaling(),  # type: ignore
        )(y)
        y = batch_normalization()(y)
        y = keras.layers.Activation(activation)(y)

        if se_ratio:
            y = se_layer(ratio=se_ratio)(y)

        y = conv2d(output_filters, kernel_size=(1, 1), strides=(1, 1))(y)
        y = batch_normalization()(y)
        y = keras.layers.Activation(activation)(y)

        if use_skip:
            x = conv2d(output_filters, kernel_size=(1, 1), strides=strides)(x)
            x = batch_normalization()(x)

        y = keras.layers.Add()([y, x])
        y = keras.layers.Activation(activation)(y)
        return y

    return layer

zblock 

zblock(output_filters: int = 0, group_width: int = 0, kernel_size: int | tuple[int, int] = 3, strides: int | tuple[int, int] = 1, se_ratio: float = 4, activation: str = 'relu6', name: str | None = None) -> Callable[[keras.KerasTensor], keras.KerasTensor]

RegNet X-Block

Parameters:

  • output_filters (int, default: 0 ) –

    Number of output filters. Defaults to 0.

  • group_width (int, default: 0 ) –

    Group width. Defaults to 0.

  • 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: 4 ) –

    SE ratio. Defaults to 4.

  • activation (str, default: 'relu6' ) –

    Activation function. Defaults to "relu6".

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

    Block name. Defaults to None.

Returns:

  • Callable[[KerasTensor], KerasTensor] –

    Callable[[keras.KerasTensor], keras.KerasTensor]: Functional layer

Source code in neuralspot_edge/models/regnet.py 
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
def zblock(
    output_filters: int = 0,
    group_width: int = 0,
    kernel_size: int | tuple[int, int] = 3,
    strides: int | tuple[int, int] = 1,
    se_ratio: float = 4,
    activation: str = "relu6",
    name: str | None = None,
) -> Callable[[keras.KerasTensor], keras.KerasTensor]:
    """RegNet X-Block

    Args:
        output_filters (int, optional): Number of output filters. Defaults to 0.
        group_width (int, optional): Group width. Defaults to 0.
        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 4.
        activation (str, optional): Activation function. Defaults to "relu6".
        name (str|None, optional): Block name. Defaults to None.

    Returns:
        Callable[[keras.KerasTensor], keras.KerasTensor]: Functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        input_filters = x.shape[-1]
        groups = input_filters // group_width
        use_add = input_filters == output_filters and (strides == 1 if isinstance(strides, int) else strides[0] == 1)
        expand_ratio = 2

        name_ex = f"{name}.exp" if name else None
        filters = input_filters * expand_ratio
        y = conv2d(filters, kernel_size=(1, 1), strides=(1, 1), name=name_ex)(x)
        y = batch_normalization(name=name_ex)(y)
        y = keras.layers.Activation(activation, name=f"{name_ex}.act")(y)

        y = keras.layers.Conv2D(
            filters,
            kernel_size=kernel_size,
            strides=strides,
            groups=groups,
            padding="same",
            use_bias=False,
            kernel_initializer=keras.initializers.VarianceScaling(),  # type: ignore
        )(y)
        y = batch_normalization()(y)
        y = keras.layers.Activation(activation)(y)

        if se_ratio:
            y = se_layer(ratio=se_ratio)(y)

        y = conv2d(output_filters, kernel_size=(1, 1), strides=(1, 1))(y)
        y = batch_normalization()(y)

        if use_add:
            y = keras.layers.Add()([y, x])

        return y

    return layer

regnet_core 

regnet_core(blocks: list[RegNetBlockParam], block_style: Literal['y', 'z'] = 'y') -> Callable[[keras.KerasTensor], keras.KerasTensor]

RegNet core

Parameters:

Returns:

  • Callable[[KerasTensor], KerasTensor] –

    Callable[[keras.KerasTensor], keras.KerasTensor]: Core

Source code in neuralspot_edge/models/regnet.py 
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
def regnet_core(
    blocks: list[RegNetBlockParam],
    block_style: Literal["y", "z"] = "y",
) -> Callable[[keras.KerasTensor], keras.KerasTensor]:
    """RegNet core

    Args:
        blocks (list[RegNetBlockParam]): Block params
        block_style (float, optional): Block style. Defaults to 'y'.

    Returns:
        Callable[[keras.KerasTensor], keras.KerasTensor]: Core
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        block_fn = yblock if block_style == "y" else zblock
        for i, block in enumerate(blocks):
            filters = make_divisible(block.filters, 8)
            for d in range(block.depth):
                name = f"stage{i}.block{d+1}"
                x = block_fn(
                    output_filters=filters,
                    group_width=block.group_width,
                    kernel_size=block.kernel_size,
                    strides=block.strides,
                    se_ratio=block.se_ratio,
                    activation=block.activation,
                    name=name,
                )(x)
            # END FOR
        # END FOR
        return x

    return layer

regnet_layer 

regnet_layer(x: keras.KerasTensor, params: RegNetParams, num_classes: int | None = None) -> keras.KerasTensor

Create RegNet TF functional model

Parameters:

  • x (KerasTensor) –

    Input tensor

  • params (RegNetParams) –

    Model parameters.

  • num_classes (int, default: None ) –

    Number of classes.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/regnet.py 
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
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
def regnet_layer(
    x: keras.KerasTensor,
    params: RegNetParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """Create RegNet TF functional model

    Args:
        x (keras.KerasTensor): Input tensor
        params (RegNetParams): 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

    # Stem
    if params.input_filters > 0:
        name = "stem"
        filters = make_divisible(params.input_filters, 8)
        y = conv2d(filters, kernel_size=(3, 3), strides=params.input_strides, name=name)(y)
        y = batch_normalization(name=name)(y)
        y = keras.layers.Activation(params.input_activation, name=f"{name}.act")(y)
    else:
        y = x

    y = regnet_core(blocks=params.blocks, block_style=params.block_style)(y)

    if params.output_filters:
        name = "neck"
        filters = make_divisible(params.output_filters, 8)
        y = conv2d(filters, kernel_size=(1, 1), strides=(1, 1), padding="valid", name=name)(y)
        y = batch_normalization(name=name)(y)
        y = keras.layers.Activation(params.output_activation, name=f"{name}.act")(y)

    if params.include_top:
        name = "top"
        y = keras.layers.GlobalAveragePooling2D(name=f"{name}.pool")(y)

        if params.dropout > 0 and 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)

    # Only reshape if needed
    elif requires_reshape:
        y = keras.layers.Reshape(y.shape[2:])(y)

    return y