dualweaver 0.1.0

DualWeaver: A PyPI package for adapting Hugging Face Time Series Foundation Models with Synergistic Feature Weaving.

DualWeaver

DualWeaver is an official PyPI package for adapting Hugging Face Time Series Foundation Models (TSFMs) using Synergistic Feature Weaving.

This package allows you to seamlessly integrate the DualWeaver methodology with any open-source pre-trained model based on the transformers library to significantly improve forecasting performance while preserving the original pre-trained knowledge.

Installation

pip install dualweaver

Or install from source:

git clone https://github.com/yourusername/dualweaver.git
cd dualweaver
pip install -e .

Quick Start

import torch
from transformers import AutoConfig, AutoModelForCausalLM
from dualweaver import DualWeaverConfig, DualWeaverModel

# 1. Load an open-source TSFM from Hugging Face
hf_config = AutoConfig.from_pretrained("thuml/timer-base-84m")
hf_model = AutoModelForCausalLM.from_config(hf_config) # or from_pretrained

# 2. Configure DualWeaver
config = DualWeaverConfig(
    adapter="WeaverMLP", # or "WeaverCNN"
    input_channel=7,
    hf_model_type="timer"
)

# 3. Wrap with DualWeaver
model = DualWeaverModel(config, hf_model)

# 4. Train
model.train()
batch_x = torch.randn(32, 672, 7) # Batch, Length, Channel
batch_y = torch.randn(32, 96, 7)  # Batch, Length, Channel
loss = model(batch_x, batch_y)
loss.backward()

# 5. Inference
model.eval()
with torch.no_grad():
    predictions = model(batch_x)

# 6. Save
model.save_pretrained("./dualweaver_model")

# 7. Load
loaded_model = DualWeaverModel.from_pretrained("./dualweaver_model")

Citation

If you find this useful or use it in your research, please consider citing the DualWeaver paper:

@misc{li2026dualweaversynergisticfeatureweaving,
      title={DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models}, 
      author={Jinpeng Li and Zhongyi Pei and Huaze Xue and Bojian Zheng and Chen Wang and Jianmin Wang},
      year={2026},
      eprint={2602.22066},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.22066}, 
}