data

MobileNetV4Config

extendsModelConfig

MobileNetV4Config(num_classes: int = 1000, in_channels: int = 3, variant: str = 'conv_small', dropout: float = 0.3, drop_path_rate: float = 0.0)

source edit

Frozen configuration for every MobileNet-v4 variant.

One dataclass describes the whole family; variant selects which of the paper's five searched architectures (Appendix D, Tables 11-15) is built, and the remaining fields cover the classifier and regularisation knobs that are not part of the searched topology.

Parameters

num_classesint= 1000

Output classes of the classification head.

in_channelsint= 3

Channels of the input image.

variantstr= "conv_small"

Which searched architecture to build — one of "conv_small", "conv_medium", "conv_large", "hybrid_medium" or "hybrid_large". The variant fixes the block sequence, the stem width (24 for the Large models, 32 otherwise), the activation (GELU for Hybrid-Large, ReLU for the rest) and whether residual branches carry a layer scale (hybrids only).

dropoutfloat= 0.3

Dropout probability before the final Linear. The paper trains Conv-Small with 0.3 and every other variant with 0.2 (Table 10); the factories set the matching value.

drop_path_ratefloat= 0.0

Peak stochastic-depth rate. Block i of n drops its residual branch with probability drop_path_rate * i / n. The paper's recipe peaks at 0 (Conv-Small), 0.075 (Medium) and 0.35 (Large) during training (Table 10); the default leaves it off so the model behaves deterministically unless a recipe asks for it.

Attributes

model_typestr

Registry id, "mobilenet_v4".

Raises

ValueError

If variant is not one of the five paper variants, or a rate lies outside

[0, 1)

Notes

Block sequences follow the authors' released implementation, which is what the published checkpoints were trained against. The four other variants match Tables 11, 12, 14 and 15 block for block. For Hybrid-Medium the released model orders several UIB blocks differently from Table 13 and carries one extra block in the stride-32 stage; its parameter count (10.56 M in units of $2^{20}$ ) is the 10.5 M that Table 6 reports.

Examples

>>> from lucid.models.vision.mobilenet_v4 import MobileNetV4Config
>>> cfg = MobileNetV4Config()
>>> cfg.model_type
'mobilenet_v4'
>>> cfg.variant, cfg.num_classes
('conv_small', 1000)
>>> MobileNetV4Config(variant="hybrid_large").variant
'hybrid_large'

Used by 3

Constructors

dunder

init

→None

__init__(num_classes: int = 1000, in_channels: int = 3, variant: str = 'conv_small', dropout: float = 0.3, drop_path_rate: float = 0.0)

source edit

>>> from lucid.models.vision.mobilenet_v4 import MobileNetV4Config >>> cfg = MobileNetV4Config() >>> cfg.model_type 'mobilenet_v4' >>> cfg.variant, cfg.num_classes ('conv_small', 1000) >>> MobileNetV4Config(variant="hybrid_large").variant 'hybrid_large'