Skip to content

mobilenetv3 with pytorch,provide pre-train model

Notifications You must be signed in to change notification settings

chenpeng68/mobilenetv3

 
 

Repository files navigation

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.

MobileNetV3

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%

About

mobilenetv3 with pytorch,provide pre-train model

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%