This is a MXNet implementation of Google's MobileNets. For details, please read the original paper:
A pretrained MobileNet model on ImageNet is provided and you can use score.py to reproduce the accuracy in ImageNet 2012 val dataset.
The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN):
| Network | Top-1 | Top-5 |
|---|---|---|
| MobileNet | 71.24 | 90.15 |
- RGB mean values [123.68,116.78,103.94] are subtracted
- scale: 0.017 is used as std values for image preprocessing
- I use ChannelwiseConvolution layer from paper
- This model is converted from MobileNet-Caffe
- If you like it, consider giving me a star.
- MXNet 11.0rc supports depthwiseConvolution now!