Command Palette
Search for a command to run...

摘要
我们介绍了下一代MobileNets,该模型基于互补搜索技术和一种新颖的架构设计。MobileNetV3通过硬件感知网络架构搜索(NAS)与NetAdapt算法相结合的方式进行了针对移动电话CPU的优化,并在此基础上通过新的架构改进进一步提升性能。本文开始探讨自动化搜索算法和网络设计如何协同工作,利用互补方法来改善整体技术水平。通过这一过程,我们创建了两个新的MobileNet模型供发布:适用于高资源使用场景的MobileNetV3-Large和适用于低资源使用场景的MobileNetV3-Small。这些模型随后被调整并应用于目标检测和语义分割任务。对于语义分割(或任何密集像素预测任务),我们提出了一种新的高效分割解码器——Lite Reduced Atrous Spatial Pyramid Pooling(LR-ASPP)。我们在移动设备上的分类、检测和分割任务中取得了最新的最佳结果。与MobileNetV2相比,MobileNetV3-Large在ImageNet分类任务上提高了3.2%的准确率,同时降低了15%的延迟;而MobileNetV3-Small则提高了4.6%的准确率,同时降低了5%的延迟。在COCO检测任务中,MobileNetV3-Large在保持与MobileNetV2相似准确率的情况下,速度提升了25%;而在Cityscapes分割任务中,MobileNetV3-Large LR-ASPP的速度比MobileNetV2 R-ASPP快30%,且准确率相当。
代码仓库
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| classification-on-indl | MobileNetV3 | Average Recall: 84.28% |
| dichotomous-image-segmentation-on-dis-te1 | MBV3 | E-measure: 0.818 HCE: 274 MAE: 0.083 S-Measure: 0.740 max F-Measure: 0.669 weighted F-measure: 0.595 |
| dichotomous-image-segmentation-on-dis-te2 | MBV3 | E-measure: 0.856 HCE: 600 MAE: 0.083 S-Measure: 0.777 max F-Measure: 0.743 weighted F-measure: 0.672 |
| dichotomous-image-segmentation-on-dis-te3 | MBV3 | E-measure: 0.880 HCE: 1136 MAE: 0.078 S-Measure: 0.764 max F-Measure: 0.772 weighted F-measure: 0.702 |
| dichotomous-image-segmentation-on-dis-te4 | MBV3 | E-measure: 0.848 HCE: 3817 MAE: 0.098 S-Measure: 0.770 max F-Measure: 0.736 weighted F-measure: 0.664 |
| dichotomous-image-segmentation-on-dis-vd | MBV3 | E-measure: 0.841 HCE: 1625 MAE: 0.092 S-Measure: 0.758 max F-Measure: 0.714 weighted F-measure: 0.642 |
| image-classification-on-imagenet | MobileNet V3-Large 1.0 | GFLOPs: 0.438 Number of params: 5.4M Top 1 Accuracy: 75.2% |
| semantic-segmentation-on-cityscapes | MobileNet V3-Large 1.0 | Mean IoU (class): 72.6% |
| semantic-segmentation-on-dada-seg | MobileNetV3 (MobileNetV3small) | mIoU: 18.2 |