Learning Robust Statistics for Simulation-Based Inference Under Model Misspecification: Code implementation

---

Note: Our implementations build on the sbi library by Macke's lab. You can find further details on its structure and the dependencies at the github website.

---

Installation

git clone https://github.com/huangdaolang/Robust-SBI.git

cd Robust-SBI

conda create --name robust-sbi python=3.8

conda activate robust-sbi

pip install -r requirements.txt

Quick Start

The description of the simulators can be found in Section 4 of our paper.

To run Ricker model, you can use the following template command lines:

python exp_ricker.py --distance="mmd" 
    --beta=1.0 
    --degree=0.0

For the above template command, we use MMD-based regularizer, and beta corresponds to the lambda in Equation 5 in our paper, which is the weight of the regularizer. The degree corresponds to the misspecification degree epsilon, where degree=0.0 means epsilon=0%.

You can also define your own simulator and corresponding summary network, which you can find more details in sbi package's official document.

If you want to skip the long simulation process, you can use generate_data.py to pre-generate all simulated data and observed data.

After training, it will return posterior, density_estimator and sum_net. For posterior and density_estimator, please check the usage in sbi package's official document. For sum_net, it is used to extract and analyse the summary statistics, you can load it with default pytorch function torch.load().

Our implementation of regularizer is in /inference/snpe/snpe_base.py, from Line 345. The summary networks can be found in inference/networks/summary_nets.py.

ABC experiments for Ricker and OUP are included in notebooks/ricker_abc.ipynb and notebooks/oup_abc.ipynb.

Code for creating all figures in our paper can be found in notebooks/plots.ipynb.

Citation

@article{huang2023learning,
  title={Learning Robust Statistics for Simulation-based Inference under Model Misspecification},
  author={Huang, Daolang and Bharti, Ayush and Souza, Amauri and Acerbi, Luigi and Kaski, Samuel},
  journal={arXiv preprint arXiv:2305.15871},
  year={2023}
}

License

This code is under the MIT License.