摘要
预训练模型逐渐成为一种工具并应用于各种下游任务。其中,由于各望远镜位置、使用数字后端、处理技术以及周围干扰环境的不同,脉冲星候选体筛选需使用极端不平衡的数据从头训练并设计深度学习模型,存在流程复杂、时效性差、难度大的问题。针对上述问题,基于ImageNet数据集预训练模型,对比分析其在脉冲星数据集的表现,并比较其Precision、Recall、F1-Score。首先,使用LogME方法评估各模型在脉冲星数据集的可迁移性;其次,使用后端融合的方式融合频率-相位图和时间-相位图的2种特征;最后,使用逐层解冻的方式微调对比可迁移性评分高的14个模型。结果表明,预训练方式所需资源少、收敛速度快,能应用于脉冲星筛选任务,在需要更进一步提升模型效果时,也可以考虑这些模型作为基础模块用于提取特征。
Pre-trained models have gradually become a tool and are applied to various downstream tasks.Among them,due to the different positions of each telescope,the use of digital backend,processing technology,and the surrounding interference environment,pulsar candidate selection requires training from scratch using extremely imbalanced data and designing deep learning models,which poses problems such as complex processes,poor timeliness,and high difficulty.In response to the above issues,a pre-trained model based on the ImageNet dataset was compared and analyzed for its performance on the pulsar dataset,including Precision,Recall,and F1 Score.Firstly,evaluate the transferability of each model on the pulsar dataset using the LogME method.Secondly,using backend fusion to fuse the two features of frequency phase and time phase maps.Finally,use a layer by layer thawing method to fine-tune the 14 models with higher transferability scores.The results show that the pre-trained method requires fewer resources and has a fast convergence speed,which can be applied to pulsar screening tasks.When further improving the model performance,these models can also be considered as basic modules for feature extraction.
作者
曾鹏
于徐红
刘志杰
王春庆
ZENG Peng;YU Xuhong;LIU Zhijie;WANG Chunqing(Key Laboratory of Information and Computing Science Guizhou Province,Guizhou Normal University,Guiyang,Guizhou 550001,China)
出处
《自动化应用》
2024年第8期227-231,共5页
Automation Application
基金
国家自然科学基金资助项目(U183110134)
中国科学院天文大科学研究中心FAST重大成果培育项目(FAST[2019sr04])。
关键词
脉冲星候选体筛选
预训练
微调
可迁移性
pulsar candidate selection
pre-trained
fine-tune
transferability
