[ad_1]
Environment friendly neural community backbones for cell gadgets are sometimes optimized for metrics comparable to FLOPs or parameter rely. Nonetheless, these metrics could not correlate nicely with latency of the community when deployed on a cell machine. Subsequently, we carry out intensive evaluation of various metrics by deploying a number of mobile-friendly networks on a cell machine. We determine and analyze architectural and optimization bottlenecks in latest environment friendly neural networks and supply methods to mitigate these bottlenecks. To this finish, we design an environment friendly spine MobileOne, with variants attaining an inference time below 1 ms on an iPhone12 with 75.9% top-1 accuracy on ImageNet. We present that MobileOne achieves state-of-the-art efficiency throughout the environment friendly architectures whereas being many occasions quicker on cell. Our greatest mannequin obtains comparable efficiency on ImageNet as MobileFormer whereas being 38× quicker. Our mannequin obtains 2.3% higher top-1 accuracy on ImageNet than EfficientNet at comparable latency. Moreover, we present that our mannequin generalizes to a number of duties – picture classification, object detection, and semantic segmentation with vital enhancements in latency and accuracy as in comparison with current environment friendly architectures when deployed on a cell machine.
Support authors and subscribe to content
This is premium stuff. Subscribe to read the entire article.
