A implementation of MobileNetV3

This is a implementation of Searching for MobileNetV3 training on pytorch, and then I convert it to caffe.

I trained the network as the paper's discription at first. The top-1 accuracy can reach 67.59% for small mobile models, but the top-1 accuracy can only reach 72.45% for large mobile models. Then I changed the order of last average pooling layer and its next convolution layer, thus the top-1 accuracy can reach 75.34% for large mobile models matching the result of the paper, and the top-1 accuracy is 68.44% for small mobile models.

Accuracy

	MAdds	Params	Top-1 acc	Pretrained Model
Offical large	219M	5.4M	75.2%	-
Offical small	66M	2.9M	67.4%	-
Ours large_old	240M	5.4M	72.%
Ours small_old	66M	2.9M	67.6%
Ours large	296M	5.4M	75.3%	google drive
Ours small	99M	2.9M	68.4%	google drive

Inference and validation

python inference.py mobilenetv3_large
python inference.py mobilenetv3_small