mobilenet_v4_conv_small

→MobileNetV41.3M

mobilenet_v4_conv_small(pretrained: bool = False, overrides: object = {})

source edit

MobileNet-v4-Conv-Small feature-extracting backbone.

Builds a MobileNetV4 with the Conv-Small architecture of Qin et al., 2024 (Appendix D, Table 11): the two stem stages are dense conv pairs, the next two are built from ExtraDW, IB and ConvNeXt-like UIBs, closed by a $1 \times 1$ widening to 960 channels. 1.26 M parameters without the classification head.

Model Size

Parameters

pretrainedbool= False

Must stay False — no checkpoint is published for the headless backbone, and True raises rather than returning random weights. The ImageNet-1k checkpoint belongs to the matching *_cls classifier, whose error message this names.

**overridesobject= {}

Keyword overrides merged into MobileNetV4Config (e.g. drop_path_rate=0.1).

Returns

MobileNetV4

Backbone returning the stride-32, 960-channel feature map.

Raises

NotImplementedError

If pretrained=True.

Notes

Qin et al., "MobileNetV4: Universal Models for the Mobile Ecosystem", ECCV 2024 (arXiv:2404.10518). The paper trains and evaluates this variant at 224x224; the network is fully convolutional up to the pool, so any input whose sides are multiples of 32 works.

Examples

>>> import lucid
>>> from lucid.models.vision.mobilenet_v4 import mobilenet_v4_conv_small
>>> model = mobilenet_v4_conv_small().eval()
>>> model(lucid.randn(1, 3, 224, 224)).last_hidden_state.shape
(1, 960, 7, 7)

Used by 1

lucid.models.vision.mobilenet_v4

>>> import lucid >>> from lucid.models.vision.mobilenet_v4 import mobilenet_v4_conv_small >>> model = mobilenet_v4_conv_small().eval() >>> model(lucid.randn(1, 3, 224, 224)).last_hidden_state.shape (1, 960, 7, 7)