Skip to content

composer

Composer Model API

This module provides utility functions to compose a sequential set of networks/layers.

Classes:

Functions:

Classes

ComposerLayerParams

Composer layer parameters

Attributes:

  • name (str) –

    Layer name

  • params (dict) –

    Layer arguments

ComposerParams

Composer Network parameters

Attributes:

  • layers (list[ComposerLayerParams]) –

    Network layers

  • include_top (bool) –

    Include top

  • output_activation (str | None) –

    Output activation

  • name (str) –

    Model name

ComposerModel

Helper class to generate model from parameters

Functions

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

Create layer from parameters

Source code in neuralspot_edge/models/composer.py 
111
112
113
114
115
116
@staticmethod
def layer_from_params(inputs: keras.Input, params: ComposerParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = ComposerParams(**params)
    return composer_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: ComposerParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/composer.py 
118
119
120
121
122
@staticmethod
def model_from_params(inputs: keras.Input, params: ComposerParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = ComposerModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

composer_layer 

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

Composes a sequential set of networks/layers. Useful for adding custom layers to a pre-trained model (e.g. foundation).

Parameters:

  • x (KerasTensor) –

    Model input

  • params (ComposerParams) –

    Model parameters

  • num_classes (int | None, default: None ) –

    Number of classes

Returns:

  • KerasTensor –

    keras.KerasTensor: Model output

Source code in neuralspot_edge/models/composer.py 
 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
def composer_layer(
    x: keras.KerasTensor,
    params: ComposerParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """Composes a sequential set of networks/layers.
    Useful for adding custom layers to a pre-trained model (e.g. foundation).

    Args:
        x (keras.KerasTensor): Model input
        params (ComposerParams): Model parameters
        num_classes (int | None): Number of classes

    Returns:
        keras.KerasTensor: Model output
    """
    y = x
    for layer in params.layers:
        match layer.name:
            case "conv2d":
                y = conv2d(y, **layer.params)
            case "dense":
                y = keras.layers.Dense(**layer.params)(y)
            case "relu6":
                y = relu6(y)
            case "batch_norm":
                y = batch_normalization(y)
            case "se_block":
                y = se_layer(y, **layer.params)
            case "load_model":
                prev_model = load_model(layer.params["model_file"])
                trainable = layer.params.get("trainable", True)
                if not trainable:
                    logger.info(f"Freezing model {prev_model.name}")
                prev_model.trainable = trainable
                y = prev_model(y, training=trainable)
            case _:
                raise ValueError(f"Unknown layer {layer.name}")
        # END MATCH
    # END FOR

    if params.include_top:
        if num_classes is not None:
            y = keras.layers.Dense(num_classes)(y)
        if params.output_activation:
            y = keras.layers.Activation(params.output_activation)(y)

    return y