SFF (Lost in the Non-convex Loss Landscape: How to Fine-tune the Large Time Series Model? Published in ICLR 2026)

March 11, 2026 · View on GitHub

Pytorch implementation of SFF. The paper is available at the link Paper (PDF).

Overview

替代文本

Datasets

The public datasets can be downloaded from https://drive.google.com/drive/folders/1PPLsAoDbv4WcoXDp-mm4LFxoKwewnKxX and place them in the datasets folder.

Usage

Timer's pre-trained weights can be downloaded from the link https://drive.google.com/drive/folders/15oaiAl4OO5gFqZMJD2lOtX2fxHbpgcU8.

In the run.py script, different evaluation modes are enabled by setting training_from_scratch (TFS), LP (linear probing), LPFF (linear probing first then full fine-tuning) or smoothed_full_finetuning. If both are set to False, the original full fine-tuning (FF) strategy is adopted.

Reference

If this repository and the work are helpful to you, please consider citing it:

@inproceedings{zhanglost,
  title={Lost in the Non-convex Loss Landscape: How to Fine-tune the Large Time Series Model?},
  author={Zhang, Xu and Wang, Peng and Wang, Wei},
  booktitle={The Fourteenth International Conference on Learning Representations}
}