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% |