neuralgbi

Project owners: Richard Gao & Michael Deistler

Amortized neural Generalized Bayesian Inference for SBI applications: using neural network-based regression and density estimation to do generalized Bayesian inference, i.e., using distance functions as pseudo-likelihood functions.

Installing dependencies

pip install -e . to run setup. pip install -e packages/sbi/ to install local version of sbi.

Generating figures

1. Run notebooks in paper/fig1/01_generate_figure.ipynb
2. Convert the svg via invoke convert 1
3. Upload to overleaf

Generating benchmark results

1. Make x_o: cd gbi/benchmark/tasks/, python generate_xo.py 'gaussian_mixture' 10
2. Generate ground-truth GBI posterior samples from x_os: cd gbi/benchmark/generate_gt/, python run_gaussian_mixture.py -m task.xo_index=0,1,2,3,4,5,6,7,8,9 task.is_specified='specified','misspecified' task.is_known='known','unknown' task.beta=2.,10.,50. task.name=gaussian_mixture
3. Train algorithms (can be done separately from step 2): cd gbi/benchmark/run_algorithms/, python run_training.py -m task.name=gaussian_mixture algorithm=NPE,NLE,GBI
4. Do inference with trained algorithms: cd gbi/benchmark/run_algorithms/, python run_inference.py -m algorithm=GBI trained_inference_datetime='$YYYY_MM_DD__hh_mm_ss' task.name='gaussian_mixture' task.xo_index=0,1,2,3,4,5,6,7,8,9 task.is_specified=specified,misspecified task.is_known=known,unknown task.beta=2.,10.,50.. Note for NPE and NLE there is no need to sweep over beta.