Machine Learning with Swift

ShuffleNet

ShuffleNet architecture was proposed in 2017 by the research team from Face++ (Megvii Inc.). It is targeted on mobile devices with limited computation power (for example, 10 - 150 MFLOPs). In comparison to the classical CNN, ShuffleNet has less parameters and performs less computations, because it uses pointwise group convolution and channel shuffle; for instance, it works 13x faster than AlexNet. The accuracy remains the same: on ImageNet, it even performs slightly better (top-1 error metric) than MobileNet.

The original paper can be found here: https://arxiv.org/abs/1707.01083.