tsnpe_neurips
收藏资源简介：
TSNPE NeurIPS 2022结果数据集是用于重现论文《Truncated proposals for scalable and hassle-free simulation-based inference》中神经科学图表的配套数据，具体涵盖幽门网络（图5）和第5层锥体细胞模型（图6）。数据集包含四个主要部分：1) l20_7/inference/目录提供第5层锥体细胞模型的两个训练运行结果，包括训练好的推理网络、验证对数概率、保留观测数据和TensorBoard日志；2) p31_4/multiround/目录包含幽门网络在多种训练配置下的多轮推理结果；3) p31_4/prior_predictives_energy_paper/目录提供用于论文能量消耗分析的先验预测模拟输出；4) simulations_pickle/目录包含第5层锥体细胞模型的模拟参数-观测对，可直接用于训练网络而无需重新运行模拟器。所有文件均为Python pickle格式，由sbi工具箱生成，适用于基于模拟的推理研究、计算神经科学和计算生物学任务。数据集规模在1K到10K之间，采用AGPL-3.0许可证。
The TSNPE NeurIPS 2022 results dataset contains training inference networks, simulation outputs, and intermediate results for reproducing the neuroscience figures (Figure 5 - pyloric network, Figure 6 - layer 5 pyramidal cell) from the paper Truncated proposals for scalable and hassle-free simulation-based inference. It is companion data for the code repository `mackelab/tsnpe_neurips`, following the `l5pc/results/` directory structure. The dataset consists of four main parts: 1) The `l20_7/inference/` directory includes two training runs for the layer 5 pyramidal cell model, containing trained inference networks, validation log probabilities, held-out observation data, and TensorBoard logs; 2) The `p31_4/multiround/` directory contains multi-round inference results for the pyloric network under various training configurations; 3) The `p31_4/prior_predictives_energy_paper/` directory provides prior predictive simulation outputs used for energy consumption analysis in the paper; 4) The `simulations_pickle/` directory includes simulated parameter-observation pairs for the layer 5 pyramidal cell model, enabling network training without re-running the simulator. All files are in Python pickle format, generated by the `sbi` toolbox, and suitable for simulation-based inference research, computational neuroscience, and computational biology tasks. The dataset size ranges from 1K to 10K and is licensed under AGPL-3.0.
- 许可证: AGPL-3.0
- 数据集名称: TSNPE NeurIPS 2022 — Results Data
- 标签: simulation-based-inference, neuroscience, computational-biology
- 规模: 1K < n < 10K
简介
该数据集包含训练后的推理网络、模拟输出和中间结果，用于复现论文《Truncated proposals for scalable and hassle-free simulation-based inference》（NeurIPS 2022）中的神经科学图表（图5——幽门网络，图6——第5层锥体细胞）。它是代码库 mackelab/tsnpe_neurips 的数据配套，镜像了该代码在 l5pc/results/ 下期望的目录结构。
内容
l20_7/inference/: 包含第5层锥体细胞（L5PC，图6）模型的两次训练运行结果：训练后的推理网络（inference.pkl）、验证对数概率、保留观测数据（xo.pkl）以及sbi训练循环中的 TensorBoard 日志（sbi-logs/）。p31_4/multiround/: 包含幽门网络（图5）在多种训练配置（带/不带变换的 APT、TSNPE）下的多轮推理结果，每个运行/轮次对应一个inference_r*.pkl文件。p31_4/prior_predictives_energy_paper/: 包含论文能量消耗分析中使用的先验预测模拟输出（all_circuit_parameters.pkl、all_simulation_outputs.pkl）。simulations_pickle/: 包含 L5PC 模型的模拟参数/观测对（simulations_theta_r*.pkl、simulations_x_r*.pkl），用于无需重新运行模拟器即可训练l20_7/inference/中的网络。
所有文件均为 Python pickle 格式，由 sbi 工具箱和配套代码库生成，可直接使用 pickle/ torch.load 加载。
使用方式
通过以下命令下载数据集： bash pip install -U "huggingface_hub[cli]" hf download mackelab/tsnpe_neurips --repo-type dataset --local-dir l5pc/results
将下载的 l5pc/results/ 目录放置在 mackelab/tsnpe_neurips 代码签出中的对应路径下，即可复现论文中的图5和图6。
相关页面
引用
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} }
