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.

Parameters are in helia_edge.models.mobileone_params.

Functions

Name	Description
`build`	MobileOne model from `MobileOneParams`
`mobileone_block`	MobileOne block
`mobileone_layer`	MobileOne layer

Additions

The MobileOne architecture has been modified to allow the following:

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

Usage

from helia_edge.models import MobileOneBlockParams, MobileOneParams, ModelSpec, build

params = MobileOneParams(
    input_filters=24,
    input_kernel_size=(1, 7),
    input_strides=(1, 2),
    input_padding=(0, 3),
    blocks=[
        MobileOneBlockParams(
            filters=32, depth=2, kernel_size=(1, 7), strides=(1, 2), padding=(0, 3), se_ratio=2, num_conv_branches=2
        )
    ],
    include_top=True,
    num_classes=5,
)
model = build(ModelSpec(params=params, input_shape=(1, 800, 1)))
model.summary()

Machine-readable model

mobileone_blockfunctionMBConv block w/ expansion and SE
mobileone_layerfunctionCreate MobileOne TF functional model
buildfunctionBuild a MobileOne model.

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

Parameters of mobileone_block
Name	Type	Default	Description
`output_filters`	`int`	Required	Number of output filter channels
`kernel_size`	`int \| tuple[int, int]`	3	Kernel size. Defaults to 3.
`strides`	`int \| tuple[int, int]`	1	Stride length. Defaults to 1.
`se_ratio`	`float`	0	SE ratio. Defaults to 8.
`name`	`str \| None`	None	Block name. Defaults to None.

Returns

Returns of mobileone_block
Type	Description
`keras.Layer`	keras.Layer: Functional layer

mobileone_layer(x: keras.KerasTensor, params: MobileOneParams, inference_mode: bool = False) -> keras.KerasTensor

Create MobileOne TF functional model

Parameters

Parameters of mobileone_layer
Name	Type	Default	Description
`x`	`keras.KerasTensor`	Required	Input tensor
`params`	`MobileOneParams`	Required	Model parameters.

Returns

Returns of mobileone_layer
Type	Description
`keras.KerasTensor`	keras.KerasTensor: Output tensor

build(
    params: MobileOneParams,
    input_shape: tuple[int | None, ...],
    *,
    batch_size: int | None = None,
    name: str | None = None,
) -> keras.Model

Build a MobileOne model.

Parameters

Parameters of build
Name	Type	Default	Description
`params`	`MobileOneParams`	Required	Model parameters.
`input_shape`	`tuple[int \| None, ...]`	Required	Input shape without the batch axis; None for a variable axis.
`batch_size`	`int \| None`	None	Static batch size; None for a dynamic batch.
`name`	`str \| None`	None	Model name; the family when None.

Returns

Returns of build
Type	Description
`keras.Model`	keras.Model: The model, named ``mobileone`` unless ``name`` is given.