tapfn_time_series for ctf4science

This submodule contains code to evaluate the foundation model TabPFNv2 (https://github.com/PriorLabs/tabpfn-time-series) on ct4science benchmarks. config_KS.yaml is a configuration file to run on all KS_Official tasks and config_Lorenz is the same for Lorenz_Official.

Usage

To better manage dependencies, we STRONGLY recommend following installation steps.

1. Create a new environment with venv or conda.
2. Pip install uv:

pip install uv

3. Install the requirements for tabpfn from the root of ctftabpfn, not the root of ctf4science.

uv pip install -r requirements.txt 

4. Install ctf4science. Make sure you are now in the root of ctf4science.

uv pip install -e .

Now, given a config file, one can train and evaluate a model by running

python run.py <path-to-config>

To generate a prediction matrix for new dataset and pair_id, follow the examples in ks_submit.py or lorenz_submit.py. For instance, running

python lorenz_submit.py --pair_id 1

will generate the prediction matrix for pair_id 1 for the Lorenz_Official data.

Citation for TabPFNv2:

@misc{hoo2025tablestimetabpfnv2outperforms, title={From Tables to Time: How TabPFN-v2 Outperforms Specialized Time Series Forecasting Models}, author={Shi Bin Hoo and Samuel Müller and David Salinas and Frank Hutter}, year={2025}, eprint={2501.02945}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2501.02945}, }