基于多任务学习的输电线路小金具缺失推理加速算法

Multi-task Learning Based Acceleration Algorithm for Missing Small Fittings in Transmission Line

摘要:

针对输电线路小金具缺失的检测问题,对小金具缺失算法的推理加速进行了研究,采用了多任务学习的方法,将小金具缺失检测任务使用一个Swin Transformer _[2^6]主干网络连接和多个MLP任务头的方式进行多任务学习和多任务推理,并进行了单任务学习和多任务学习的精度和性能对比实验,最后还验证了在多任务学习中无缝增加扩展任务,实验结果表明多任务学习的输电线路小金具缺失推理在比单任务学习的推理速度提升了2倍以上,在推理显存占用上降低了22%以上。通过扩展任务实验结果验证了扩展任务的有效性,提高了任务配置的灵活性。

Abstract:

Aiming at the detection problem of missing small fittings in transmission lines, the acceleration of the inference speed of the missing small fittings algorithm is studied. The multi-task learning method is adopted, and the missing small fittings detection task is connected by a Swin Transformer backbone network and multiple MLP task heads. We propose a new approach of doing multi-task learning as well as multi-task inference, and compare the accuracy against single-task learning in different scenarios. We then verified that new tasks can be seamlessly added to the multi-task learning framework, without any performance degradation. The experimental results show that the speed of proposed multi-task learning is more than 2 times higher than that of single-task learning, and the memory usage is reduced by more than 22%. These experimental results verified the effectiveness of the proposed multi-task learning framework and the flexibility of the task settings.