Skip to content

resnet

ResNet

Overview

ResNet is a type of convolutional neural network (CNN) that is commonly used for image classification tasks. ResNet is a fully convolutional network that consists of a series of convolutional layers and pooling layers. The pooling layers are used to downsample the input while the convolutional layers are used to upsample the input. The skip connections between the pooling layers and convolutional layers allow ResNet to preserve spatial/temporal information while also allowing for faster training and inference times.

For more info, refer to the original paper Deep Residual Learning for Image Recognition.

Classes:

Functions:

Additions

  • Enable 1D and 2D variants.

Classes

ResNetBlockParams

ResNet 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

  • bottleneck (bool) –

    Use bottleneck blocks

  • activation (str) –

    Activation function

ResNetParams

ResNet parameters

Attributes:

  • blocks (list[ResNetBlockParams]) –

    ResNet blocks

  • input_filters (int) –

    Input filters

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

    Input kernel size

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

    Input stride

  • input_activation (str) –

    Input activation

  • include_top (bool) –

    Include top

  • output_activation (str | None) –

    Output activation

  • dropout (float) –

    Dropout rate

  • name (str) –

    Model name

ResNetModel

Helper class to generate model from parameters

Functions

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

Create layer from parameters

Source code in neuralspot_edge/models/resnet.py 
226
227
228
229
230
231
@staticmethod
def layer_from_params(inputs: keras.Input, params: ResNetParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = ResNetParams(**params)
    return resnet_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: ResNetParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/resnet.py 
233
234
235
236
237
@staticmethod
def model_from_params(inputs: keras.Input, params: ResNetParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = ResNetModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

generate_bottleneck_block 

generate_bottleneck_block(filters: int, kernel_size: int | tuple[int, int] = 3, strides: int | tuple[int, int] = 1, expansion: int = 4, activation: str = 'relu6') -> keras.Layer

Generate functional bottleneck block.

Parameters:

  • filters (int) –

    Filter size

  • 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.

  • expansion (int, default: 4 ) –

    Expansion factor. Defaults to 4.

Returns:

  • Layer –

    keras.Layer: TF functional layer

Source code in neuralspot_edge/models/resnet.py 
 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
def generate_bottleneck_block(
    filters: int,
    kernel_size: int | tuple[int, int] = 3,
    strides: int | tuple[int, int] = 1,
    expansion: int = 4,
    activation: str = "relu6",
) -> keras.Layer:
    """Generate functional bottleneck block.

    Args:
        filters (int): Filter size
        kernel_size (int | tuple[int, int], optional): Kernel size. Defaults to 3.
        strides (int | tuple[int, int], optional): Stride length. Defaults to 1.
        expansion (int, optional): Expansion factor. Defaults to 4.

    Returns:
        keras.Layer: TF functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        num_chan = x.shape[-1]
        projection = num_chan != filters * expansion or (strides > 1 if isinstance(strides, int) else strides[0] > 1)

        bx = conv2d(filters, 1, 1)(x)
        bx = batch_normalization()(bx)
        bx = keras.layers.Activation(activation)(bx)

        bx = conv2d(filters, kernel_size, strides)(x)
        bx = batch_normalization()(bx)
        bx = keras.layers.Activation(activation)(bx)

        bx = conv2d(filters * expansion, 1, 1)(bx)
        bx = batch_normalization()(bx)

        if projection:
            x = conv2d(filters * expansion, 1, strides)(x)
            x = batch_normalization()(x)
        x = keras.layers.Add()([bx, x])
        x = keras.layers.Activation(activation)(x)
        return x

    return layer

generate_residual_block 

generate_residual_block(filters: int, kernel_size: int | tuple[int, int] = 3, strides: int | tuple[int, int] = 1, activation: str = 'relu6') -> keras.Layer

Generate functional residual block

Parameters:

  • filters (int) –

    Filter size

  • 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.

Returns:

  • Layer –

    keras.Layer: TF functional layer

Source code in neuralspot_edge/models/resnet.py 
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
def generate_residual_block(
    filters: int,
    kernel_size: int | tuple[int, int] = 3,
    strides: int | tuple[int, int] = 1,
    activation: str = "relu6",
) -> keras.Layer:
    """Generate functional residual block

    Args:
        filters (int): Filter size
        kernel_size (int | tuple[int, int], optional): Kernel size. Defaults to 3.
        strides (int | tuple[int, int], optional): Stride length. Defaults to 1.

    Returns:
        keras.Layer: TF functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        num_chan = x.shape[-1]
        projection = num_chan != filters or (strides > 1 if isinstance(strides, int) else strides[0] > 1)
        bx = conv2d(filters, kernel_size, strides)(x)
        bx = batch_normalization()(bx)
        bx = keras.layers.Activation(activation)(bx)

        bx = conv2d(filters, kernel_size, 1)(bx)
        bx = batch_normalization()(bx)
        if projection:
            x = conv2d(filters, 1, strides)(x)
            x = batch_normalization()(x)
        x = keras.layers.Add()([bx, x])
        x = keras.layers.Activation(activation)(x)
        return x

    return layer

resnet_layer 

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

Generate functional ResNet model. Args: x (keras.KerasTensor): Inputs params (ResNetParams): Model parameters. num_classes (int, optional): Number of class outputs. Defaults to None.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/resnet.py 
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
def resnet_layer(
    x: keras.KerasTensor,
    params: ResNetParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """Generate functional ResNet model.
    Args:
        x (keras.KerasTensor): Inputs
        params (ResNetParams): Model parameters.
        num_classes (int, optional): Number of class outputs. Defaults to None.

    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

    if params.input_filters:
        y = conv2d(
            params.input_filters,
            kernel_size=params.input_kernel_size,
            strides=params.input_strides,
        )(y)
        y = batch_normalization()(y)
        y = keras.layers.Activation(params.input_activation)(y)
    # END IF

    for stage, block in enumerate(params.blocks):
        for d in range(block.depth):
            func = generate_bottleneck_block if block.bottleneck else generate_residual_block
            y = func(
                filters=block.filters,
                kernel_size=block.kernel_size,
                strides=block.strides if d == 0 and stage > 0 else 1,
                activation=block.activation,
            )(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)

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

    return y