预训练模型的跨领域跨任务迁移学习

为了解决计算机深度学习时标注数据工作量大、准确度不高、耗时耗力等问题,需要将预先训练好的模型中的数据进行跨领域跨任务迁移学习。基于对不同数据集的迁移效果的研究,试验时将视觉领域中表现良好的ImageNet预训练模型迁移到音频分类任务,通过剔除无声部分、统一音频长度、数据转换和正则化处理这4个步骤,采用经典图像增强和样本混淆两种数据增强方法,以5种不同的方式训练数据集,实验证明:ImageNET目标训练模型的跨领域迁移学习效果显著,但源领域的模型效果和目标领域的最终效果并没有必然联系,且使用同领域相似数据分布的数据集的预训练效果比ImageNet上的预训练效果更差。

In order to solve the problems of large workloads,low accuracy and time-consuming in data-labeling in deep learning,it is necessary to transfer the data from the pre-trained model to cross-domain/cross-task learning.Based on the study of the migration effect of different dat sets,the ImageNet pre-training model,which is good in the visual f ield,is migrated t o t he audio classif ication t ask.By eliminating t he s ilent part,unifying t he audio length,data conversion and regularization processing,classical image enhancement and sample confusion are used to enhance the data,f ive different training methods to train datasets.Experiments show that ImageNET target training model has signif icant effect on cross-domain migration learning,but the effect of source domain model is not necessarily related to the f inal effect of target domain,and the pre-training effect of datasets with similar data distribution in the same f ield is worse than that on ImageNet.