MobileNetV4 - Universal Models for the Mobile Ecosystem

2024年04月16日
  • 简介
    我们推出了最新一代MobileNets,称为MobileNetV4(MNv4),为移动设备提供了通用高效的架构设计。在其核心,我们介绍了通用倒置瓶颈(UIB)搜索块,这是一个统一而灵活的结构,将倒置瓶颈(IB)、ConvNext、前馈网络(FFN)和一种新颖的Extra Depthwise(ExtraDW)变体合并在一起。除了UIB,我们还提出了Mobile MQA,这是一个专门为移动加速器量身定制的注意力块,可提供39%的加速。我们还介绍了一种优化的神经架构搜索(NAS)配方,提高了MNv4搜索的有效性。UIB、Mobile MQA和优化后的NAS配方的整合,导致了一系列MNv4模型的推出,这些模型在移动CPU、DSP、GPU以及专用加速器(如Apple Neural Engine和Google Pixel EdgeTPU)上大多数是帕累托最优的,这是其他测试模型所没有的特点。最后,为了进一步提高准确性,我们介绍了一种新颖的蒸馏技术。通过这种技术的增强,我们的MNv4-Hybrid-Large模型在Pixel 8 EdgeTPU运行时仅需3.8ms,提供了87%的ImageNet-1K准确性。
  • 图表
  • 解决问题
    MobileNetV4: A Universal Architecture for Mobile Devices
  • 关键思路
    The paper introduces the Universal Inverted Bottleneck (UIB) search block, a unified and flexible structure that merges Inverted Bottleneck (IB), ConvNext, Feed Forward Network (FFN), and a novel Extra Depthwise (ExtraDW) variant. Alongside UIB, Mobile MQA, an attention block tailored for mobile accelerators, delivering a significant 39% speedup. An optimized neural architecture search (NAS) recipe is also introduced which improves MNv4 search effectiveness.
  • 其它亮点
    The new suite of MNv4 models are mostly Pareto optimal across mobile CPUs, DSPs, GPUs, as well as specialized accelerators like Apple Neural Engine and Google Pixel EdgeTPU. The MNv4-Hybrid-Large model delivers 87% ImageNet-1K accuracy, with a Pixel 8 EdgeTPU runtime of just 3.8ms. The paper also introduces a novel distillation technique.
  • 相关研究
    Some related studies in this field include EfficientNet, MnasNet, and MobileNetV3.
PDF
原文
点赞 收藏 评论 分享到Link

沙发等你来抢

去评论