MobileNetV3 Backbone

A feature extractor based on MobileNetV3, refactored from https://github.com/xiaolai-sqlai/mobilenetv3: backbone only — the classification head is removed, with unified naming.

Models

Model	Variant	Output features	Weight file
MobileNetV3	small / large (auto)	1280	detected from checkpoint
MobileNetV3_Small	Small	1280	mobilenetv3_small.safetensors
MobileNetV3_Large	Large	1280	mobilenetv3_large.safetensors

Input: (B, 3, H, W) images. Output: (B, 1280) feature vectors (migrated from ImageNet-pretrained weights).

Weight naming

The naming differences between the old and the new code are mapped automatically during migration:

Old name	New name
bn1/bn2/bn3	norm1/norm2/norm3
linear3	proj
Block.se.se.*	Block.se.features.*
linear4 (classification head)	removed

Usage

from mobile_net import MobileNetV3_Small, MobileNetV3_Large

fe = MobileNetV3_Large().load_pretrained('./mobilenetv3_large.safetensors')
fe.eval()

fe.save_pretrained('./backbone.safetensors')
fe.save_pretrained('./backbone.pth')

import torch
x = torch.randn(2, 3, 224, 224)
feat = fe(x) # (2, 1280)

Auto backend detection

MobileNetV3.from_pretrained inspects the checkpoint and picks Small / Large automatically:

from mobile_net import MobileNetV3

fe = MobileNetV3.from_pretrained('./mobilenetv3_large.safetensors')  # -> Large, weights loaded
print(fe.backend)   # 'large'

Calling it on a pinned class (MobileNetV3_Large / MobileNetV3_Small) raises if the checkpoint backend disagrees with the class.