README.md

---
license: agpl-3.0
pretty_name: TSNPE NeurIPS 2022 - Results Data
tags:
- simulation-based-inference
- neuroscience
- computational-biology
size_categories:
- 1K<n<10K
---

# TSNPE NeurIPS 2022 — Results Data

This dataset contains the trained inference networks, simulation outputs, and
intermediate results used to reproduce the neuroscience figures (Fig. 5 —
pyloric network, Fig. 6 — layer 5 pyramidal cell) of:

> Michael Deistler, Pedro J. Gonçalves, Jakob H. Macke.
> **Truncated proposals for scalable and hassle-free simulation-based inference.**
> NeurIPS 2022. https://openreview.net/forum?id=QW98XBAqNRa

It is the data companion to the code repository
[`mackelab/tsnpe_neurips`](https://github.com/mackelab/tsnpe_neurips), and mirrors
the directory structure that code expects under `l5pc/results/`. These files were
previously stored via Git LFS directly in that repository; they now live here to
keep the code repo's clone size and Git LFS bandwidth low.

## Contents

- **`l20_7/inference/`** — two training runs for the layer 5 pyramidal cell (L5PC,
  Fig. 6) model: trained inference networks (`inference.pkl`), validation
  log-probabilities, held-out observations (`xo.pkl`), and TensorBoard logs
  (`sbi-logs/`) from the `sbi` training loop.
- **`p31_4/multiround/`** — multi-round inference results for the pyloric network
  (Fig. 5) across several training configurations (APT with/without transformation,
  TSNPE), one `inference_r*.pkl` per run/round.
- **`p31_4/prior_predictives_energy_paper/`** — prior predictive simulation outputs
  (`all_circuit_parameters.pkl`, `all_simulation_outputs.pkl`) used for the energy
  consumption analysis in the paper.
- **`simulations_pickle/`** — simulated parameter/observation pairs
  (`simulations_theta_r*.pkl`, `simulations_x_r*.pkl`) for the L5PC model, used to
  train the networks in `l20_7/inference/` without re-running the simulator.

All files are Python pickles produced by the `sbi` toolbox and the code in the
companion repository; there is no `datasets`-library loading script — load them
directly with `pickle`/`torch.load` as consumed by the notebooks in
[`mackelab/tsnpe_neurips`](https://github.com/mackelab/tsnpe_neurips).

## Usage

```bash
pip install -U "huggingface_hub[cli]"
hf download mackelab/tsnpe_neurips --repo-type dataset --local-dir l5pc/results
```

Place the downloaded `l5pc/results/` directory at the corresponding path inside a
checkout of [`mackelab/tsnpe_neurips`](https://github.com/mackelab/tsnpe_neurips)
to reproduce Fig. 5 and Fig. 6 of the paper.

## Citation

```bibtex
@inproceedings{
  deistler2022truncated,
  title={Truncated proposals for scalable and hassle-free simulation-based inference},
  author={Michael Deistler and Pedro J. Goncalves and Jakob H. Macke},
  booktitle={Thirty-Sixth Conference on Neural Information Processing Systems},
  year={2022},
  url={https://openreview.net/forum?id=QW98XBAqNRa}
}
```

## License

AGPL-3.0, matching the [code repository](https://github.com/mackelab/tsnpe_neurips/blob/main/LICENSE).

## Contact

If you have questions, please reach out to `michael.deistler@uni-tuebingen.de`.