We just test four models in ImageNet-1K, both train set and val set are scaled to 256(minimal side), only use Mirror and RandomResizeCrop as training data augmentation, during validation, we use center crop to get 224x224 patch.
CPU Info: Intel(R) Xeon(R) CPU E5-2650 v3 @ 2.30GHz
ImageNet-1K
Models
validation(Top-1)
validation(Top-5)
CPU Cost(ms)
MnasNet
64.91
86.28
~300
ShuffleNetV2 x1
61.83
83.99
~100
Note
Maybe the implement of these network have some different from origin method, we can not achieve the best performance as said in the paper.