Created by: shippingwang
This PR
- add efficientnet-lite model
There are some changes between EfficientNet and EfficientNet-Lite model,
one purpose is trying to suit the edge device,
- delete se
- local_pooling
- fix head and stem (reduce params)
the other is better for post-quantization.
- swish -> relu6
- bilinear intepolation
- inception preprocessing
performance: