标题： 用于心房颤动分类的ResNet架构基准测试 显示英文标题

标题： Benchmarks of ResNet Architecture for Atrial Fibrillation Classification

摘要： 在本工作中，我们将ResNet架构的不同变体应用于心房颤动分类任务。 不同之处在于第一个卷积后的滤波器数量、ResNet块布局、块卷积中的滤波器数量以及下采样操作之间的ResNet块数量。 我们发现了一个模型规模范围，在此范围内，配置差异较大的模型表现出相似的性能。 总体参数数量可能在模型性能中起主导作用。 然而，像布局这样的配置参数始终能带来更好的结果，这表明这些参数应首先被定义和固定，而其他参数可以在合理范围内变化以满足任何现有约束。 显示英文摘要

摘要： In this work we apply variations of ResNet architecture to the task of atrial fibrillation classification. Variations differ in number of filter after first convolution, ResNet block layout, number of filters in block convolutions and number of ResNet blocks between downsampling operations. We have found a range of model size in which models with quite different configurations show similar performance. It is likely that overall number of parameters plays dominant role in model performance. However, configuration parameters like layout have values that constantly lead to better results, which allows to suggest that these parameters should be defined and fixed in the first place, while others may be varied in a reasonable range to satisfy any existing constraints.