MobileNetV5Backbone model

[source]

MobileNetV5Backbone class

keras_hub.models.MobileNetV5Backbone(
    stackwise_block_types,
    stackwise_num_blocks,
    stackwise_num_filters,
    stackwise_strides,
    stackwise_act_layers,
    stackwise_exp_ratios,
    stackwise_se_ratios,
    stackwise_dw_kernel_sizes,
    stackwise_dw_start_kernel_sizes,
    stackwise_dw_end_kernel_sizes,
    stackwise_exp_kernel_sizes,
    stackwise_pw_kernel_sizes,
    stackwise_num_heads,
    stackwise_key_dims,
    stackwise_value_dims,
    stackwise_kv_strides,
    stackwise_use_cpe,
    filters=3,
    stem_size=16,
    stem_bias=True,
    fix_stem=False,
    num_features=2048,
    pad_type="same",
    use_msfa=True,
    msfa_indices=(-2, -1),
    msfa_output_resolution=16,
    act_layer="gelu",
    norm_layer="rms_norm",
    se_layer=keras_hub.src.models.mobilenet.mobilenet_backbone.SqueezeAndExcite2D,
    se_from_exp=True,
    round_chs_fn=round_channels,
    drop_path_rate=0.0,
    layer_scale_init_value=None,
    image_shape=(None, None, 3),
    data_format=None,
    dtype=None,
    **kwargs
)

MobileNetV5 backbone network.

This class represents the backbone of the MobileNetV5 architecture, which can be used as a feature extractor for various downstream tasks.

Arguments

• stackwise_block_types: list of list of strings. The block type for each block in each stack.
• stackwise_num_blocks: list of ints. The number of blocks for each stack.
• stackwise_num_filters: list of list of ints. The number of filters for each block in each stack.
• stackwise_strides: list of list of ints. The stride for each block in each stack.
• stackwise_act_layers: list of list of strings. The activation function for each block in each stack.
• stackwise_exp_ratios: list of list of floats. The expansion ratio for each block in each stack.
• stackwise_se_ratios: list of list of floats. The SE ratio for each block in each stack.
• stackwise_dw_kernel_sizes: list of list of ints. The depthwise kernel size for each block in each stack.
• stackwise_dw_start_kernel_sizes: list of list of ints. The start depthwise kernel size for each uir block in each stack.
• stackwise_dw_end_kernel_sizes: list of list of ints. The end depthwise kernel size for each uir block in each stack.
• stackwise_exp_kernel_sizes: list of list of ints. The expansion kernel size for each er block in each stack.
• stackwise_pw_kernel_sizes: list of list of ints. The pointwise kernel size for each er block in each stack.
• stackwise_num_heads: list of list of ints. The number of heads for each mqa or mha block in each stack.
• stackwise_key_dims: list of list of ints. The key dimension for each mqa or mha block in each stack.
• stackwise_value_dims: list of list of ints. The value dimension for each mqa or mha block in each stack.
• stackwise_kv_strides: list of list of ints. The key-value stride for each mqa or mha block in each stack.
• stackwise_use_cpe: list of list of bools. Whether to use conditional position encoding for each mqa or mha block in each stack.
• filters: int. The number of input channels.
• stem_size: int. The number of channels in the stem convolution.
• stem_bias: bool. If True, a bias term is used in the stem convolution.
• fix_stem: bool. If True, the stem size is not rounded.
• num_features: int. The number of output features, used when use_msfa is True.
• pad_type: str. The padding type for convolutions.
• use_msfa: bool. If True, the Multi-Scale Fusion Adapter is used.
• msfa_indices: tuple. The indices of the feature maps to be used by the MSFA.
• msfa_output_resolution: int. The output resolution of the MSFA.
• act_layer: str. The activation function to use.
• norm_layer: str. The normalization layer to use.
• se_layer: keras.layers.Layer. The Squeeze-and-Excitation layer to use.
• se_from_exp: bool. If True, SE channel reduction is based on the expanded channels.
• round_chs_fn: callable. A function to round the number of channels.
• drop_path_rate: float. The stochastic depth rate.
• layer_scale_init_value: float. The initial value for layer scale.
• image_shape: tuple. The shape of the input image. Defaults to (None, None, 3).
• data_format: str, The data format of the image channels. Can be either "channels_first" or "channels_last". If None is specified, it will use the image_data_format value found in your Keras config file at ~/.keras/keras.json. Defaults to None.
• dtype: None or str or keras.mixed_precision.DTypePolicy. The dtype to use for the model's computations and weights. Defaults to None.

Example

import keras
from keras_hub.models import MobileNetV5Backbone

# Randomly initialized backbone with a custom config.
model_args = {
    "stackwise_block_types": [["er"], ["uir", "uir"]],
    "stackwise_num_blocks": [1, 2],
    "stackwise_num_filters": [[24], [48, 48]],
    "stackwise_strides": [[2], [2, 1]],
    "stackwise_act_layers": [["relu"], ["relu", "relu"]],
    "stackwise_exp_ratios": [[4.0], [6.0, 6.0]],
    "stackwise_se_ratios": [[0.0], [0.0, 0.0]],
    "stackwise_dw_kernel_sizes": [[0], [5, 5]],
    "stackwise_dw_start_kernel_sizes": [[0], [0, 0]],
    "stackwise_dw_end_kernel_sizes": [[0], [0, 0]],
    "stackwise_exp_kernel_sizes": [[3], [0, 0]],
    "stackwise_pw_kernel_sizes": [[1], [0, 0]],
    "stackwise_num_heads": [[0], [0, 0]],
    "stackwise_key_dims": [[0], [0, 0]],
    "stackwise_value_dims": [[0], [0, 0]],
    "stackwise_kv_strides": [[0], [0, 0]],
    "stackwise_use_cpe": [[False], [False, False]],
    "use_msfa": False,
}
model = MobileNetV5Backbone(**model_args)
input_data = keras.ops.ones((1, 224, 224, 3))
output = model(input_data)

# Load the backbone from a preset and run a prediction.
backbone = MobileNetV5Backbone.from_preset("mobilenetv5_300m_gemma3n")

# Expected output shape = (1, 16, 16, 2048).
outputs = backbone.predict(keras.ops.ones((1, 224, 224, 3)))

[source]

from_preset method

MobileNetV5Backbone.from_preset(preset, load_weights=True, **kwargs)

Instantiate a keras_hub.models.Backbone from a model preset.

A preset is a directory of configs, weights and other file assets used to save and load a pre-trained model. The preset can be passed as a one of:

1. a built-in preset identifier like 'bert_base_en'
2. a Kaggle Models handle like 'kaggle://user/bert/keras/bert_base_en'
3. a Hugging Face handle like 'hf://user/bert_base_en'
4. a ModelScope handle like 'modelscope://user/bert_base_en'
5. a path to a local preset directory like './bert_base_en'

This constructor can be called in one of two ways. Either from the base class like keras_hub.models.Backbone.from_preset(), or from a model class like keras_hub.models.GemmaBackbone.from_preset(). If calling from the base class, the subclass of the returning object will be inferred from the config in the preset directory.

For any Backbone subclass, you can run cls.presets.keys() to list all built-in presets available on the class.

Arguments

• preset: string. A built-in preset identifier, a Kaggle Models handle, a Hugging Face handle, or a path to a local directory.
• load_weights: bool. If True, the weights will be loaded into the model architecture. If False, the weights will be randomly initialized.

Examples

# Load a Gemma backbone with pre-trained weights.
model = keras_hub.models.Backbone.from_preset(
    "gemma_2b_en",
)

# Load a Bert backbone with a pre-trained config and random weights.
model = keras_hub.models.Backbone.from_preset(
    "bert_base_en",
    load_weights=False,
)