Informer2020 无阉割分步复现:官方 Jupyter 版拆解为 .py 工程,ProbSparse Attention + 生成式解码,ETTh1 滑窗预测(seq96 / label48 / pred24),「6 轮早停 vs 30 轮不早停」对照实验定量验证早停价值(测试 R² 0.77 vs 0.45),踩坑与调参笔记全程留痕。
python machine-learning deep-learning time-series pytorch transformer forecasting reproducibility informer time-series-forecasting long-sequence etth1 probsparse-attention
-
Updated
Jan 22, 2026 - Python