For ShuffleNetV2, the speed of inference in cpu is slow

vision
1

Sorry, I encountered a problem. When I trained the model “shuffleNetv2” on GPU, I found it that the speed of training was very fast, having obvious advantages than other CNN models. However, the speed of inference on cpu was slower than that of other CNN models obviously, thus, I would know is there some problems in my experiments? Please give me some solutions if possible? Thanks a lot!

2

Did you experience the same thing with ShuffleNet v1?
ShuffleNet have splitting layers, which split the input to two parallel process.
Additionally, ShuffleNetv2 was design to be GPU efficient and ARM architecture. I wonder it would hurt the CPU speed.

3

Thanks a lot! I will perform the same operation with ShuffleNetV1. And, if using ShuffleNets on intel cpu, the speed cannot give any improvements? And try other lightweight models, only?