This is a Gluon implementation of Mnasnet architecture as described in the paper MnasNet: Platform-Aware Neural Architecture Search for Mobile.
To train this model, you need first install the gluon-cv or download the repo beside as we do. Using train_imagenet.sh
to train, you will get 0.73+ accuracy.
A pretrained model is provided by zuesees. They've got 73.6% Top-1 score and 91.52% Top-5 score, nearly approaching official accuracy.
After some tunes, I think you will get higher accuracy. I'm not free to do this now, so I hope someone could report to me when he/she get a better accuracy.