ddz16/preformer

This repository contains the pytorch code for the 2023 ICASSP paper "Preformer: Predictive Transformer with Multi-Scale Segment-wise Correlations for Long-Term Time Series Forecasting”

Hi 👋, I'm ddz

• 📚 I am interested in time series forecasting & video understanding. In terms of video understanding, I am focusing on intelligent processing for underwater untrimmed long videos.

• 😀 You can follow me on Zhihu.

• 📧 How to reach me [email protected].

  Universities and Institutions

  Tech Stack

  Git Stats

preformer's Issues

RuntimeError: shape '[128, 1, 1]' is invalid for input of size 0

Are you the first author of this paper？

An issue about the multi-scale hybrid decomposition

作者您好，请问论文里提到的多尺度混合分解部分体现在代码里的哪个部分？不清楚是不是我理解有误，在/models/Preformer.py/20、21
kernel_size = configs.moving_avg
self.decomp = series_decomp(kernel_size)
中只看到了一种指定尺度的分解？Preformer中的series_decomp和Autoformer源代码里的series_decomp有哪些不同呢？

Q和K的计算方式不是哈达玛乘积

在selfattention_family.py中的SegmentCorrelation中，correlation_scores = torch.einsum("bmlhd,bnlhd->bhmn", seg_queries, seg_keys)本质上不还是点积运算吗