- An MXNet implementation of mobilenetv2. Inverted Residuals and Linear Bottlenecks Mobile Networks for Classification Detection and Segmentation.
- use relu as it activation layer rather than use relu6 in inverted residual bottleneck layers.
We got 0.235 map on coco datasets for mobilenetv2+ssdlite with 512 input size,here. Imagenet model will be released as soon as we reach the accuacy of the paper.
We tested our code on:
Ubuntu 16.04, Python 2.7 with
numpy(1.11.0), cv2(3.3.0-dev)
mxnet 0.11.0