A PyTorch implementation of MobileNetV3

I make a mistake to forget the avgpool in se model, now I have re-trained the mbv3_small, the mbv3_large is on training, it will be coming soon.

You should use torch.load to load the model.

model = torch.load("mbv3_large.old.pth.tar", map_location='cpu')
weight = model["state_dict"]

This is a PyTorch implementation of MobileNetV3 architecture as described in the paper Searching for MobileNetV3.

Some details may be different from the original paper, welcome to discuss and help me figure it out.


Madds Parameters Top1-acc
Large 219 M 5.4 M 75.2%
Small 66 M 2.9 M 67.4%
Ours Large old 272 M 3.96 M 75.454%
Ours Small old 66 M 2.51 M 69.069%
Ours Large new 265 M 3.96 M
Ours Small new 64 M 2.51 M 69.037%


🚀 Github 镜像仓库 🚀




贡献者 2


  • Python 100.0 %