Thinklab-SJTU / Crossformer

Official implementation of our ICLR 2023 paper "Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting"
Apache License 2.0
476 stars 84 forks source link

关于MTGNN #21

Open teethdocker opened 8 months ago

teethdocker commented 8 months ago

您好,我有问题想向您请教,我看到文中实验加入了MTGNN的比较,有比较好的效果,但我对MTGNN源代码进行改写后进行实验出现了OOM问题,请问您是如何处理这个问题的呢?

YunhaoZhang-Mars commented 8 months ago

在模型结构上减小hidden state的数量,训练时用较小的batch size;另外还需要将原实现中的loss改为MSE loss。