AggCVAE/AggCVAE_thesis_github

Deep Learning and MCMC Inference for Mapping HIV Prevalence in Zambia. # AggCVAE_thesis_github Deep learning and MCMC inference for mapping HIV Prevalence in Zambia 2018. The code in this repository builds on the work from ES: github.com. This repository is organised as follows: - `aggGP.ipynb`: code for the AggGP model. - `aggVAE_indiv.ipynb`: code for the AggVAE model applied to administrative units 1 and 2 separately. - `aggCVAE_gpu_indiv_admin.ipynb`: code for the AggCVAE model applied to administrative units 1 and 2 separately. - `aggCVAE_all_admin.ipynb`: code for the AggCVAE model encoding administrative units 1 and 2 jointly. - `aggCVAE_thesis_figures.ipynb`: code to produce comparison plots between aggVAE, aggCVAE and aggGP in the thesis. The code relies heavily on the `numpyro` probabilistic programming library for Bayesian hierarchical modelling and MCMC inference, as well as `JAX` for deep learning. It is recommended that the reader familiarises themselves with the corresponding documentation before using the code: - `numpyro`: num.pyro.ai - `JAX`: jax.readthedocs.io For the models, each notebook follows a similar structure: 1. Import libraries 2. Set the administrative level (if applicable) 3. Read shapefiles of boundaries of the administrative units 4. Read HIV prevalence data 5. Merge the shape and prevalence data 6. Create the computational grid - Create a regular grid - Add points to administrative units with insufficient gridpoints 8. Train a VAE/CVAE algorithm with `JAX` (if applicable) - Define preliminary functions (covariance matrix of GP, indicator matrix M) - Define the model architecture (model configurations, MLP layers, etc.) - Define the aggregated GP prior to be approximated by the VAE/CVAE with `numpyro` - Define functions needed for training (dataloader, training step, validation step, loss function) - Train the VAE/CVAE - Save the decoder and losses - Evaluate how well the VAE/CVAE has learnt from …