The paper "ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design" by Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun from Megvii Inc (Face++) and Tsinghua University, addresses the limitations of evaluating neural network architecture design solely based on computation complexity metrics like FLOPs. The authors propose practical guidelines for efficient network design, emphasizing the importance of considering direct metrics such as speed and memory access cost (MAC) over indirect metrics like FLOPs. They conduct a series of controlled experiments to derive these guidelines and present a new architecture, ShuffleNet V2, which is designed to optimize speed and accuracy. Key findings and guidelines include: 1. **Equal Channel Width**: Using equal input and output channels minimizes MAC. 2. **Group Convolution**: Excessive group convolution increases MAC. 3. **Network Fragmentation**: Reducing network fragmentation improves parallelism. 4. **Element-wise Operations**: These operations are non-negligible and should be minimized. ShuffleNet V2 incorporates these guidelines by introducing a channel split operation, which helps maintain a large number of equally wide channels while reducing MAC. The architecture is tested on ImageNet for classification and COCO for object detection, demonstrating superior performance in terms of speed and accuracy compared to other state-of-the-art models like ShuffleNet v1, MobileNet v2, and Xception. The paper also discusses the generalization of ShuffleNet V2 to large models and its effectiveness in object detection tasks.The paper "ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design" by Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun from Megvii Inc (Face++) and Tsinghua University, addresses the limitations of evaluating neural network architecture design solely based on computation complexity metrics like FLOPs. The authors propose practical guidelines for efficient network design, emphasizing the importance of considering direct metrics such as speed and memory access cost (MAC) over indirect metrics like FLOPs. They conduct a series of controlled experiments to derive these guidelines and present a new architecture, ShuffleNet V2, which is designed to optimize speed and accuracy. Key findings and guidelines include: 1. **Equal Channel Width**: Using equal input and output channels minimizes MAC. 2. **Group Convolution**: Excessive group convolution increases MAC. 3. **Network Fragmentation**: Reducing network fragmentation improves parallelism. 4. **Element-wise Operations**: These operations are non-negligible and should be minimized. ShuffleNet V2 incorporates these guidelines by introducing a channel split operation, which helps maintain a large number of equally wide channels while reducing MAC. The architecture is tested on ImageNet for classification and COCO for object detection, demonstrating superior performance in terms of speed and accuracy compared to other state-of-the-art models like ShuffleNet v1, MobileNet v2, and Xception. The paper also discusses the generalization of ShuffleNet V2 to large models and its effectiveness in object detection tasks.