CupidShuffleNet
abstract
With the recent popularity of Vision Transformers, much of the focus has been on the novel aspect of the field. This paper looks to simply expand on one of the influential edge-focused models, The ShuffleNet. We propose CupidShuffle, a ShuffleNetV2 variant replacing the initial entry layer with a Window-based Transformer and also removing 7 ShuffleV2Unit layers while still keeping the round-trip shuffle intact by switching from [3, 7, 3] to [1, 4, 1]. The proposed architecture keeps ShuffleNetV2’s predictive power, while decreasing the total model size. And probably the most important contribution, this paper also attempts to keep it short and concise.