DualWeaver: Synergistic Feature Fusion Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models

This repository contains the implementation for our ICML submission.

1. Environment Setup

To set up the environment, please install the required dependencies using pip. We recommend using a virtual environment.

pip install -r requirements.txt

2. Apply Custom Patch to Transformers

To ensure correct backpropagation for the AdaPTS, a patched version of transformers.generation.utils.py is required.

1. Locate your installed transformers package directory:

   pip show transformers

   (Look for the Location path in the output, e.g., /path/to/site-packages)

2. Overwrite the original generation/utils.py with the provided patch:

   # Example:
   cp patch/transformers.generation.utils.py /path/to/site-packages/transformers/generation/utils.py

3. Model Preparation

Please ensure the TSFM checkpoints are placed correctly. Download the pre-trained timer-base-84m (or other supported models) from Hugging Face and place the model.safetensors and other config files into the hf_ltm directory.

Structure should look like:

hf_ltm/
└── timer-base-84m/
    ├── config.json
    ├── model.safetensors
    └── ...

4. Run Example

We provide a script to run a demonstration experiment on the ETTh1 dataset using TimerXL and WeaverMLP.

bash scripts/example.sh