We provide pretrained MobileNetV2(width_mult: 1.0, crop_size: 224) models on ImageNet, which achieve slightly better accuracy rates than the original ones reported in the paper.
The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN):
Network
Top-1
Top-5
MobileNet v2
71.72
90.13
Training Script
For training, please refer to liangfu/mxnet-mobilenet-v2. Our pretrained model is converted from that. The MXNet officially offer a pretrained model (and training script) here too.
Normalization
The input images are substrated by mean RGB = [ 123.68, 116.78, 103.94 ].