biubiu0719 / mobilenetv3

mobilenetv3 with pytorch，provide pre-train model

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%