efficientnet

EfficientNetV2

Overview

EfficientNetV2 is an improvement to EfficientNet that incorporates additional optimizations to reduce both computation and memory. In particular, the architecture leverages both fused and non-fused MBConv blocks, non-uniform layer scaling, and training-aware NAS.

For more info, refer to the original paper EfficientNetV2: Smaller Models and Faster Training.

Classes:

• EfficientNetV2Params –

  EfficientNetV2 parameters

• EfficientNetV2Model –

  Helper class to generate model from parameters

Functions:

• efficientnetv2_layer –

  EfficientNetV2 layer

Additions

The EfficientNetV2 architecture has been modified to allow the following:

• Enable 1D and 2D variants.

Usage

import keras
from neuralspot_edge.models import EfficientNetV2, EfficientNetV2Params, MBConvParams

inputs = keras.Input(shape=(800, 1))
num_classes = 5

model = EfficientNetV2(
    x=inputs,
    params=EfficientNetV2Params(
        input_filters=24,
        input_kernel_size=(1, 7),
        input_strides=(1, 2),
        blocks=[
            MBConvParams(filters=32, depth=2, kernel_size=(1, 7), strides=(1, 2), ex_ratio=1, se_ratio=2),
            MBConvParams(filters=48, depth=2, kernel_size=(1, 7), strides=(1, 2), ex_ratio=1, se_ratio=2),
            MBConvParams(filters=64, depth=2, kernel_size=(1, 7), strides=(1, 2), ex_ratio=1, se_ratio=2),
            MBConvParams(filters=72, depth=1, kernel_size=(1, 7), strides=(1, 2), ex_ratio=1, se_ratio=2)
        ],
        output_filters=0,
        include_top=True,
        dropout=0.2,
        drop_connect_rate=0.2,
        model_name="efficientnetv2"
    ),
    num_classes=num_classes,
)

Classes

EfficientNetParams

EfficientNet parameters

Attributes:

• blocks (list[MBConvParams]) –

  EfficientNet 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

• output_filters (int) –

  Output filters

• output_activation (str | None) –

  Output activation

• include_top (bool) –

  Include top

• dropout (float) –

  Dropout rate

• drop_connect_rate (float) –

  Drop connect rate

• use_logits (bool) –

  Use logits

• activation (str) –

  Activation function

• norm (Literal['batch', 'layer'] | None) –

  Normalization type

• name (str) –

  Model name

EfficientNetV2Model

Helper class to generate model from parameters

Functions

layer_from_params staticmethod

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

Create layer from parameters

Source code in neuralspot_edge/models/efficientnet.py

@staticmethod
def layer_from_params(inputs: keras.Input, params: EfficientNetParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = EfficientNetParams(**params)
    return efficientnetv2_layer(x=inputs, params=params, num_classes=num_classes)

model_from_params staticmethod

model_from_params(inputs: keras.Input, params: EfficientNetParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/efficientnet.py

@staticmethod
def model_from_params(inputs: keras.Input, params: EfficientNetParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = EfficientNetV2Model.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

efficientnet_core

efficientnet_core(blocks: list[MBConvParams], drop_connect_rate: float = 0) -> keras.Layer

EfficientNet core

Parameters:

• blocks

  (list[MBConvParam]) –

  MBConv params

• drop_connect_rate

  (float, default: 0 ) –

  Drop connect rate. Defaults to 0.

Returns:

• Layer –

  keras.Layer: Core

Source code in neuralspot_edge/models/efficientnet.py

def efficientnet_core(blocks: list[MBConvParams], drop_connect_rate: float = 0) -> keras.Layer:
    """EfficientNet core

    Args:
        blocks (list[MBConvParam]): MBConv params
        drop_connect_rate (float, optional): Drop connect rate. Defaults to 0.

    Returns:
        keras.Layer: Core
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        global_block_id = 0
        total_blocks = sum((b.depth for b in blocks))
        for i, block in enumerate(blocks):
            filters = make_divisible(block.filters, 8)
            for d in range(block.depth):
                name = f"stage{i+1}_mbconv{d+1}"
                block_drop_rate = drop_connect_rate * global_block_id / total_blocks
                x = mbconv_block(
                    filters,
                    block.ex_ratio,
                    block.kernel_size,
                    block.strides if d == 0 else 1,
                    block.se_ratio,
                    droprate=block_drop_rate,
                    bn_momentum=block.bn_momentum,
                    activation=block.activation,
                    name=name,
                )(x)
                global_block_id += 1
            # END FOR
        # END FOR
        return x

    # END DEF
    return layer

efficientnetv2_layer

efficientnetv2_layer(x: keras.KerasTensor, params: EfficientNetParams, num_classes: int | None = None) -> keras.KerasTensor

Create EfficientNet V2 TF functional model

Parameters:

• x

  (KerasTensor) –

  Input tensor

• params

  (EfficientNetParams) –

  Model parameters.

• num_classes

  (int, default: None ) –

  Number of classes.

Returns:

• KerasTensor –

  keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/efficientnet.py

def efficientnetv2_layer(
    x: keras.KerasTensor,
    params: EfficientNetParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """Create EfficientNet V2 TF functional model

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

    Returns:
        keras.KerasTensor: Output tensor
    """

    # Force input to be 4D (add dummy dimension)
    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=params.input_kernel_size,
            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)
    # END IF

    y = efficientnet_core(blocks=params.blocks, drop_connect_rate=params.drop_connect_rate)(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="same", 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 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)
        elif not params.use_logits:
            y = keras.layers.Softmax()(y)
    # Only reshape if required
    elif requires_reshape:
        y = keras.layers.Reshape(y.shape[2:])(y)

    return y